WEBVTT
00:00:00.959 --> 00:00:08.640
You know, the theme behind my work is using data and analytical tools to understand complex problems.
00:00:08.640 --> 00:00:18.719
Try to find patterns in data, understanding why these patterns exist, and what insights that we can glean from this, you know, for practice.
00:00:19.039 --> 00:00:26.000
We live in a data-rich environment these days, but translating data into truly useful information is challenging for business leaders.
00:00:26.000 --> 00:00:29.199
How can we tame data to gain valuable insights?
00:00:29.199 --> 00:00:36.880
We explore this issue and more in this episode of Inside Alvarez Business, a podcast dedicated to bringing you the stories behind the research.
00:00:36.880 --> 00:00:43.840
I'm your host, Jonathon Halbesleben, Dean of the Carlos Alvarez College of Business at The University of Texas at San Antonio.
00:00:43.840 --> 00:00:45.759
Our guest today is Saidat Sanni.
00:00:45.759 --> 00:00:53.920
We had a terrific conversation about how she has developed new approaches to extrapolate information from massive data sets to produce powerful research insights.
00:00:53.920 --> 00:00:57.200
Enjoy this episode of Inside Alvarez Business.
00:00:57.200 --> 00:01:04.879
Today I am joined by Saidat Sanni, assistant professor in the Department of Operations and Analytics in the Carlos Alvarez College of Business.
00:01:04.879 --> 00:01:07.920
She's also an Alvarez Research Fellow here in the college.
00:01:07.920 --> 00:01:15.680
She completed her PhD in business analytics at the University of Cincinnati in 2024 and then joined us shortly after that.
00:01:15.680 --> 00:01:20.000
She's still relatively new in her career, but is already making a pretty huge impact.
00:01:20.000 --> 00:01:26.400
She's recently had papers accepted in some of our leading journals, including Biometrics and Information Systems Research.
00:01:26.400 --> 00:01:34.480
Her work spans disciplinary boundaries, but a common theme is developing machine learning and statistical models for high-dimensional data.
00:01:34.480 --> 00:01:36.640
She's joined us at a perfect time.
00:01:36.640 --> 00:01:40.640
Of course, it would have been a perfect time no matter when you joined us, but you joined us at a perfect time.
00:01:40.640 --> 00:01:46.560
We're launching a new master's in science and business analytics, and so that'll be very helpful to have your expertise as we join that.
00:01:46.560 --> 00:01:49.680
So, Saidat, welcome to Inside Alvarez Business.
00:01:50.000 --> 00:01:51.200
Thank you so much for having me.
00:01:51.200 --> 00:01:53.519
I'm excited to be here and happy to chat.
00:01:53.760 --> 00:01:54.000
Great.
00:01:54.000 --> 00:01:56.560
Well, and I'm looking forward to hearing some of the stories behind your research.
00:01:56.560 --> 00:01:57.840
So this is - this is perfect.
00:01:58.079 --> 00:01:58.560
Yes.
00:01:58.799 --> 00:02:01.120
Well, let's start kind of way at the beginning.
00:02:01.120 --> 00:02:08.879
As frequent listeners to the podcast, which both of them, both frequent listeners, will know.
00:02:08.879 --> 00:02:12.960
I always ask people like why they decided to become a professor.
00:02:12.960 --> 00:02:17.439
Because I don't think there's anybody that grew up saying, I'm going to be a professor of business analytics when I grow up.
00:02:17.439 --> 00:02:20.560
So for you, what - what was the kind of driving factor?
00:02:20.560 --> 00:02:23.520
What was the moment when you decided you wanted to do this for a living?
00:02:23.840 --> 00:02:24.240
Right.
00:02:24.240 --> 00:02:29.520
I think I never, you know, as a kid, I didn't say I wanted to be a professor.
00:02:29.520 --> 00:02:39.439
But I think it just is a combination of things that I naturally enjoy, even though I didn't know earlier on that it was gonna be a career path for me.
00:02:39.439 --> 00:02:46.800
I'm originally from Nigeria, and education is an important part of being Nigerian, you know.
00:02:46.800 --> 00:02:50.719
In my family, it just felt like that's the natural path.
00:02:50.719 --> 00:02:55.680
You know, you get a degree and you know, you just keep going as much as you can.
00:02:55.680 --> 00:03:05.280
And also given that I'm the first child of my family at a young age, I would usually help them with assignments.
00:03:05.280 --> 00:03:07.520
I would tutor my siblings during the holidays.
00:03:07.520 --> 00:03:11.039
So they're busy and not just, you know, they're busy with work.
00:03:11.039 --> 00:03:13.840
You know, my dad would always have me, have me do that.
00:03:13.840 --> 00:03:15.360
And I enjoyed it.
00:03:15.360 --> 00:03:17.439
I went on to high school.
00:03:17.439 --> 00:03:18.960
I continued to do that.
00:03:18.960 --> 00:03:24.479
I enjoyed explaining difficult problems to people, you know, kind of break it down so it's easier.
00:03:24.639 --> 00:03:24.800
Yeah.
00:03:24.960 --> 00:03:30.080
I realized that it also helped me retain the knowledge or the concept.
00:03:30.080 --> 00:03:32.719
Went on to college, continued doing the same thing.
00:03:32.719 --> 00:03:35.680
Even at the time, it wasn't, it just felt like, oh, I enjoyed doing this.
00:03:35.680 --> 00:03:36.719
Oh, can you teach me this?
00:03:36.719 --> 00:03:37.680
Oh, of course I will.
00:03:37.680 --> 00:03:46.240
In terms of then, you know, deciding to be a professor, really started, I started really thinking about that when I started research.
00:03:46.240 --> 00:03:51.039
During my master's program, I'd left Nigeria and moved to the United States for that.
00:03:51.039 --> 00:03:53.039
I started getting into research.
00:03:53.039 --> 00:04:11.199
I enjoyed the idea of having questions and being able to kind of walk through these questions that don't have obvious answers, and then being able to walk through the process of doing the research, solving this problem, and eventually being able to garner some insights from this.
00:04:11.199 --> 00:04:12.560
So I really did like that.
00:04:12.560 --> 00:04:16.959
And then I decided, hey, I'll just get a PhD, you know, might as well.
00:04:16.959 --> 00:04:18.399
I continued research.
00:04:18.399 --> 00:04:22.959
And, you know, in summary, it just kind of fell into place for me.
00:04:22.959 --> 00:04:23.920
I enjoyed teaching.
00:04:23.920 --> 00:04:25.519
Now I enjoy research.
00:04:25.519 --> 00:04:31.040
These two things, going to the career of being a professor, you know, it just felt natural.
00:04:31.600 --> 00:04:32.000
Oh, good.
00:04:32.000 --> 00:04:32.720
Oh, good.
00:04:32.720 --> 00:04:34.240
What a great story.
00:04:34.240 --> 00:04:35.920
So I'm, I'm always curious.
00:04:35.920 --> 00:04:41.600
You so you grew up in Nigeria, and at some point you you made that decision to come to the United States for your master's degree.
00:04:41.759 --> 00:04:41.920
Yes.
00:04:42.160 --> 00:04:43.759
What kind of led to that decision?
00:04:44.079 --> 00:04:44.800
Oh, right.
00:04:44.800 --> 00:04:55.199
Again, it goes back to the culture that my family, or I would say broadly, Nigeria has with always seeking for knowledge.
00:04:55.199 --> 00:05:00.800
You know, after I got the bachelor's degree, I wanted to do something more.
00:05:00.800 --> 00:05:03.680
I felt that, okay, now I have the basics in statistics.
00:05:03.680 --> 00:05:05.759
I have an undergraduate in statistics.
00:05:05.759 --> 00:05:08.240
I can see the practicality of statistics.
00:05:08.240 --> 00:05:10.240
Can I take this a notch farther?
00:05:10.240 --> 00:05:12.639
And then I decided I like the practice.
00:05:12.639 --> 00:05:14.240
How about I get an MBA?
00:05:14.240 --> 00:05:16.000
And so I came to the U.S.
00:05:16.000 --> 00:05:17.519
to get an MBA.
00:05:17.519 --> 00:05:24.560
Really, I just felt, you know, it's the opportunity of being able to advance in that line since I enjoyed this so much.
00:05:24.560 --> 00:05:31.839
Even though, you know, I had people who have supported me along the way, but I also had people who said, this was their opinion, not mine.
00:05:31.839 --> 00:05:34.720
Oh, statistics, it's very technical.
00:05:34.720 --> 00:05:36.399
How do you do business?
00:05:36.399 --> 00:05:41.920
Because then you have to write papers, you have to do presentations and things like that.
00:05:41.920 --> 00:05:44.959
And I just felt, well, that's the practice.
00:05:44.959 --> 00:05:48.319
That's what made me fall in love with statistics in the first place.
00:05:48.319 --> 00:05:51.279
And so that was why I made the decision to move here.
00:05:51.279 --> 00:05:59.360
And after I had the business degree, I decided statistics plus business, let's get a PhD in business analytics.
00:05:59.360 --> 00:06:00.879
And that's how I got there.
00:06:01.120 --> 00:06:01.920
Oh, very good.
00:06:01.920 --> 00:06:02.639
Very good.
00:06:02.639 --> 00:06:04.160
So I, I'm curious then.
00:06:04.160 --> 00:06:13.279
I this might kind of, you sort of answered this question for me, but kind of more specific in terms of the research that you've now pursued since then.
00:06:13.279 --> 00:06:17.360
What, what inspired you to study some of the topics that you've studied?
00:06:17.360 --> 00:06:26.000
We'll talk about a couple of your papers in specific in a little bit here, but just kind of generally speaking, like what got you into some of the areas that you've, you've studied now.
00:06:26.399 --> 00:06:26.959
Right.
00:06:26.959 --> 00:06:38.240
The motivation behind, you know, the theme behind my work is using data and analytical tools to understand complex problems.
00:06:38.240 --> 00:06:48.959
You know, try to find patterns in data, understanding why these patterns exist and what insights that we can glean from this, you know, for practice.
00:06:48.959 --> 00:07:07.360
And majority of my work, while some are methodological, focusing on developing statistical tools to handle complex data, you would also find that I also study technology and human behavior, also driven from you know, looking at complex data sets as well.
00:07:07.360 --> 00:07:09.839
My application then spans in different areas.
00:07:09.839 --> 00:07:27.759
It could be we're going from genetics, which is you know, developing statistical methodology to financial markets, you know, putting forward models to predict stock and bond returns, and of course, digital platforms as well, understanding how technology and interpersonal relationships interplay on online platforms.
00:07:28.160 --> 00:07:36.800
I'm interested in your work in part because I had a mentor kind of early in my career when I was my first job was in a med school, and I didn't know anything about medicine.
00:07:36.800 --> 00:07:38.639
Kind of a problem.
00:07:38.639 --> 00:07:41.439
So I did a lot of work studying in a hospital.
00:07:41.439 --> 00:07:49.519
And one of the things that always stuck with me is I had this mentor who said, This hospital has got tons and tons of data, and we have so little information.
00:07:49.519 --> 00:07:56.959
And so I think you know, folks like you that are kind of taking these huge data sets and trying to distill like what's, what's the real information there?
00:07:56.959 --> 00:07:58.639
I mean, that's, that's - there's a lot of value in that.
00:07:58.639 --> 00:08:01.839
So I'm, I'm excited that we've got you working on some of those things.
00:08:01.839 --> 00:08:06.000
So let's talk about some of your work in a little bit more detail.
00:08:06.000 --> 00:08:11.439
So I'm gonna start with the paper that you've got forthcoming in the journal of Information Systems Research.
00:08:11.439 --> 00:08:12.639
I thought this one was really cool.
00:08:12.639 --> 00:08:33.840
I mean, I thought both papers were talking about were really cool, but ah, so this one it studied how racial similarity between Airbnb hosts and the guests that stay in their properties, of course, how that similarity, the racial similarity impacted the reviews of the guests, which is really kind of an intriguing area.
00:08:33.840 --> 00:08:36.720
I had certainly, I didn't really, hadn't really thought much about that.
00:08:36.720 --> 00:08:56.240
So you had predicted that guests in marginalized pairs, so two like two members of marginalized groups that are together in a pair, they would have more positive reviews, and that then there were some features of the technology that might actually increase that even further, where they have more or more positive reviews, if that's a if that's a phrase.
00:08:56.240 --> 00:09:03.600
So what - I know we were talking about, what inspired you to study these things, but that that's a pretty specific thing to study (laughter).
00:09:03.600 --> 00:09:07.440
So what inspired that kind of really specific thing to take a look at?
00:09:07.759 --> 00:09:08.320
Right.
00:09:08.320 --> 00:09:24.399
Well, digital platforms, you know, be it Airbnb, Uber, or platforms where you have people trying to borrow money, borrowers and lenders, you have, you know, platforms for service providers and you know, people that are trying to use the service, kind of, you know, it's commonplace.
00:09:24.399 --> 00:09:25.600
It's pretty much everywhere.
00:09:25.600 --> 00:09:31.200
And you have people who are then connected on those platforms and they conduct transactions.
00:09:31.200 --> 00:09:46.320
Studies, and even we humans might believe that, you know, if people are similar, you know, they have similar interests, similar lived experiences, similar values, they might likely have some form of familiarity, which could actually, you know, create some form of trust.
00:09:46.320 --> 00:09:54.399
Because to kind of do transactions on those platforms, you need some level of trust to conduct such, such transactions.
00:09:54.639 --> 00:09:56.159
Yeah, I mean you're staying in somebody's home.
00:09:56.480 --> 00:09:56.799
Exactly.
00:09:56.799 --> 00:10:12.720
Like Airbnb, you're staying in somebody's home, someone who you don't have a relationship, you don't know their history, but based on some information that you gleaned from this platform, or maybe if you've had prior conversations with them online on this digital platform, you might, you know, take a chance, you know, doing that.
00:10:12.720 --> 00:10:28.399
And so what really drove the motivation was besides that digital platforms are commonplace, it's to kind of not only say that because people are similar, they have similar interests, then it could lead to perceived positive experience with this transaction or in this digital platform.
00:10:28.399 --> 00:10:30.639
Rather, we're looking at the nuances.
00:10:30.639 --> 00:10:32.639
Maybe similarity is not enough.
00:10:32.639 --> 00:10:44.960
Maybe there's some features that this digital platform has put in place, however simple, however technical, but they have put in place to kind of help this interaction between them.
00:10:44.960 --> 00:10:59.200
And with these technology features, we could see that, you know, similarity could lead to benefits in terms of online, you know, passivity, perceived experience, positive perceived experiences on these platforms.
00:10:59.200 --> 00:11:02.559
But really, what we're looking at is maybe similarity is not enough.
00:11:02.559 --> 00:11:09.279
Maybe there's something else that adds on to that that then allows us to see the benefits of being similar on that platform.
00:11:09.279 --> 00:11:19.600
And so, for example, in Airbnb, you're trying to lodge in somebody's home, you barely know them, but you're conducting this interaction online or this transaction online.
00:11:19.600 --> 00:11:26.879
You go ahead and lodge in their place, something happens probably offline, and then you come back online to kind of give a review.
00:11:26.879 --> 00:11:29.679
The start and end of this relationship happens online.
00:11:29.679 --> 00:11:33.440
So technology is a medium with which these interactions happen.
00:11:33.519 --> 00:11:33.679
Yeah.
00:11:33.919 --> 00:11:46.240
And what we're trying to kind of look at here is how similarity and digital platforms shape perceived experiences on online, pretty much through online reviews.
00:11:46.639 --> 00:11:46.879
Okay.
00:11:46.879 --> 00:11:48.000
And it was interesting.
00:11:48.000 --> 00:11:55.440
So the some of the technology features you looked at were things like the instant booking and like flexible cancellation policies.
00:11:55.440 --> 00:12:05.679
And so it was it was kind of interesting to me that those were features you thought would actually sort of enhance the the effect of the similarity.
00:12:05.679 --> 00:12:11.519
What what led you to that idea that it would be those features that, that would be so important in this?
00:12:12.000 --> 00:12:12.399
Right.
00:12:12.399 --> 00:12:14.559
What we what led to that?
00:12:14.559 --> 00:12:17.039
Well, okay.
00:12:17.039 --> 00:12:23.440
So we found we had, you know, we're thinking, now we're saying going beyond similarity.
00:12:23.440 --> 00:12:33.440
But when you look at similarity at first, even before we go beyond this, we're saying maybe similarities in interests, similarity and lived experiences, similarities and values.
00:12:33.440 --> 00:12:40.399
And we chose this feature specifically because with cancellation, you're looking at flexibility.
00:12:40.559 --> 00:12:40.879
Yeah.
00:12:41.759 --> 00:12:52.159
You have a host who is willing to extend some flexibility to the guest, and a guest who probably likes the idea of, oh, I can cancel at any time, you know, if this arrangement doesn't work for me anymore.
00:12:52.159 --> 00:12:55.519
That's something that it probably may have in common.
00:12:55.679 --> 00:12:55.919
Yeah.
00:12:56.159 --> 00:12:56.559
Right?
00:12:56.559 --> 00:13:00.559
Instant booking, it's we look at it from the lens of shared control.
00:13:00.559 --> 00:13:06.639
What I mean by shared control is that as the host, you have a lot more information about this unit.
00:13:06.639 --> 00:13:08.000
You have more information.
00:13:08.000 --> 00:13:18.720
So if we look at more information as saying you have more control, right, you're also the one who can some of these features of the transaction are controlled by you.
00:13:18.720 --> 00:13:19.759
Who do I accept?
00:13:19.759 --> 00:13:21.120
Who do I not accept?
00:13:21.120 --> 00:13:37.120
By allowing instant booking, this means that with some pre-specified criteria, the guest can actually book to stay in this list in this lodging without having to go through, you know, sending your request and all that.
00:13:37.120 --> 00:13:38.399
It's pretty much instant.
00:13:38.480 --> 00:13:38.639
Right.
00:13:38.799 --> 00:13:42.000
So that way you're sharing some form of control with this person.
00:13:42.000 --> 00:13:46.960
You're allowing that person to have some kind of say-so in this transaction that's going on.
00:13:47.120 --> 00:13:47.279
Okay.
00:13:47.440 --> 00:13:55.840
And so we are chosen this specifically to see that, hey, not only are these people similar in some form, because in terms of similarity, we're looking at social identity here.
00:13:55.840 --> 00:14:12.080
If these two people happen to be similar in that way, if there's now an added feature of, hey, I'm offering some flexibility, I'm offering some shared control, do we see then that the benefits of similarity is weakened or strengthened?
00:14:12.159 --> 00:14:12.320
Yeah.
00:14:12.559 --> 00:14:18.480
And that was really what draw interest in that, you know, something like the host is also giving something that the guest wants.
00:14:18.799 --> 00:14:19.519
Right, right.
00:14:19.519 --> 00:14:32.720
I always I find it interesting that the cancellation policy one I found particularly interesting, mostly because if, if you get to the point where you're leaving a review, you presumably didn't cancel the reservation.
00:14:32.720 --> 00:14:35.840
So it feels a little like almost counterintuitive.
00:14:35.840 --> 00:14:38.399
Like, why would that be involved in this process?
00:14:38.399 --> 00:14:44.960
But I do, I can see that that notion of kind of the shared value of flexibility, and there's some control there too.
00:14:44.960 --> 00:14:46.639
Like I can control if I need to cancel.
00:14:46.639 --> 00:14:49.279
And so that that one was kind of interesting to me.
00:14:49.279 --> 00:14:52.159
Just it felt counterintuitive, but I get it.
00:14:52.159 --> 00:14:53.039
I get it now.
00:14:53.039 --> 00:14:55.519
So thank you for helping me understand.
00:14:55.519 --> 00:15:06.080
So what was interesting, kind of speaking of counterintuitive, you found a little bit of a counterintuitive finding in the sense that you'd expected just the similarity to lead to higher reviews.
00:15:06.080 --> 00:15:15.679
And that wasn't necessarily the case, but once you looked at the features, then the similarity, you know, then there was that effect of higher reviews.
00:15:15.840 --> 00:15:16.000
Right.
00:15:16.240 --> 00:15:17.600
What do you suppose that was?
00:15:18.080 --> 00:15:27.840
I, I guess it just goes back to because people are similar, does not necessarily mean that the interaction is perceived as a positive experience.
00:15:28.080 --> 00:15:28.320
Okay.
00:15:28.559 --> 00:15:33.279
Even though, you know, data platforms try to bring people together, connect them together, which is great.
00:15:33.279 --> 00:15:37.600
But what we are saying here is we should look beyond similarity.
00:15:37.600 --> 00:15:39.440
Similarity is not enough.
00:15:39.600 --> 00:15:39.759
Yeah.
00:15:39.919 --> 00:15:43.200
We need to put forward features.
00:15:43.200 --> 00:15:59.919
Might be simple, but simple features that govern this interaction, such that, you know, it bolsters similarity, but at the same time, it allows people to kind of - the technology aligns with their values, things they appreciate.
00:15:59.919 --> 00:16:04.399
Not to think that, you know, technology is going to have an even effect for everybody.
00:16:04.559 --> 00:16:04.720
Yeah.
00:16:04.879 --> 00:16:07.840
That's really what we're, you know, part of what we're looking at.
00:16:07.840 --> 00:16:11.360
But again, I would say similarity is not enough.
00:16:11.360 --> 00:16:18.240
There are features that govern this transaction such that, you know, both parties kind of enjoy.
00:16:18.240 --> 00:16:24.799
In our case, we're looking at, you know, in our case, we're looking at the guests' experience after the leave a review.
00:16:24.879 --> 00:16:25.200
Right.
00:16:25.360 --> 00:16:43.840
But again, we're saying that this positive reviews that we see are more likely, you know, not just because of similarity, but it's more likely because there's an interplay within, there's an interplay of similarity and the technological features that lead to, you know, positive experiences.
00:16:44.240 --> 00:16:44.559
Okay.
00:16:44.559 --> 00:16:46.320
Interesting.
00:16:46.320 --> 00:16:55.600
One of the things that I think I found really interesting was not just like what you found, but the way you went about finding it.
00:16:55.600 --> 00:17:02.799
So I think one of the things that's maybe hard for a listener to appreciate if they haven't stumbled upon your paper otherwise.
00:17:02.799 --> 00:17:07.200
It's just the massive amount of data you're talking about.
00:17:07.200 --> 00:17:18.000
So you, you were looking at between two cities, New Orleans and New York, over 50,000 listings and like close to 1.5 million reviews.
00:17:18.000 --> 00:17:22.400
And it's not as though you like read through 1.5 million reviews.
00:17:22.400 --> 00:17:25.039
So they like they had a good time.
00:17:25.039 --> 00:17:28.240
So clearly, that's a lot of data to sort through.
00:17:28.240 --> 00:17:31.200
Can you talk a little bit about how you tackled that issue?
00:17:31.200 --> 00:17:33.920
Because I think that's one of the kind of defining features of the study.
00:17:34.240 --> 00:17:38.000
Yes, this was a lot of data, and these were texts.
00:17:38.000 --> 00:17:40.400
These were online reviews that were left by guests.
00:17:40.480 --> 00:17:40.799
Yeah.
00:17:40.960 --> 00:17:47.920
And, you know, yes, the numbers, we had a huge data to work through, huge data to, you know, pre-process.
00:17:47.920 --> 00:17:55.200
Eventually, we in the paper, eventually our findings were based on 500,000 online reviews.
00:17:55.200 --> 00:18:05.039
And what we had done here was apply a transformer-based AI model, distilBERT model, to get the sentiment in these reviews.
00:18:05.039 --> 00:18:13.839
Now we went beyond just counting the number of positive words and negative words to see how positive or how negative this review or each review is.
00:18:13.839 --> 00:18:17.920
What we had done is this transformer-based model is pre-trained.
00:18:17.920 --> 00:18:25.920
And so we're able to see the probability that this is a positive review, neutral, or probability that it's a negative review in terms of sentiment.
00:18:26.079 --> 00:18:26.240
Yeah.
00:18:26.480 --> 00:18:28.880
And, you know, that's really what we applied.
00:18:28.880 --> 00:18:36.960
And then the sentiment then became our outcome measure, our response variable, with which we are now trying to see, hey, what is the effect of similarity on review sentiment?
00:18:36.960 --> 00:18:41.839
What is the effect of, you know, use of this platform features?
00:18:41.839 --> 00:18:53.359
You know, when we, besides similarity, when we look at the interaction of similarity and some of these features, instant booking, a cancellation policy, do we see that reviews are more likely to be positive or otherwise?
00:18:53.359 --> 00:19:05.039
And what we did find is, you know, what we did find from this study is that, again, when we look at the interaction with these features, instant booking, then we find that the reviews are more likely to be positive.
00:19:05.039 --> 00:19:06.720
So it goes beyond just being similar.
00:19:06.720 --> 00:19:24.799
But as an add-on, when you have this policies or not policies, when you have these features that allow for some form of flexibility, that have some form of shared control, that have that aligns more with this us and guest experiences, then we tend to see more positive reviews.
00:19:25.119 --> 00:19:25.359
Yeah.
00:19:25.359 --> 00:19:26.559
So I'm really curious.
00:19:26.559 --> 00:19:31.920
This is gonna demonstrate my, my lack of understanding in some ways.
00:19:31.920 --> 00:19:38.240
So I'm thinking, you know, you've got this machine learning model that's reading through these reviews.
00:19:38.240 --> 00:19:44.960
Presumably it's trained to sort of understand context and almost like slang.
00:19:44.960 --> 00:19:53.200
I mean, I'm thinking when I talk to my kids and I'll say something that's, you know, they'll they'll say something that's really great and they'll be like, yeah, that's fire.
00:19:53.200 --> 00:19:56.480
That would be terrible in an Airbnb review.
00:19:56.720 --> 00:19:57.359
Right?
00:19:58.079 --> 00:20:04.000
You know, or I had a killer time, you know, like that'd be a really problematic in an Airbnb review.
00:20:04.000 --> 00:20:18.160
If, if it the context wasn't kind of there, so the models are they're trained to sort of understand these types of nuances in human communication, or is there, I guess, how do you train the model to understand those nuances?
00:20:18.480 --> 00:20:18.720
Right.
00:20:18.720 --> 00:20:20.400
That that's an excellent question.
00:20:20.400 --> 00:20:25.359
I think that a lot of models have advanced, you know, there've been a lot of advancement.
00:20:25.359 --> 00:20:33.839
Earlier, you know, when you when you're looking at review sentiment, you could count the number of positive words.
00:20:33.839 --> 00:20:37.200
So you could do one word at a time, pretty much dictionary-based.
00:20:37.440 --> 00:20:37.839
Right, right.
00:20:37.839 --> 00:20:38.400
Right.
00:20:38.640 --> 00:20:41.119
And then, you know, that loses the context.
00:20:41.119 --> 00:20:47.759
Because then, you know, when you break up words, you lose the meaning, the the meaning that it gives as an entire sentence.
00:20:47.920 --> 00:20:48.319
Right, right.
00:20:48.480 --> 00:20:48.799
Right?
00:20:48.799 --> 00:20:51.440
But a lot of models have advanced beyond that now.
00:20:51.440 --> 00:20:57.680
Like the model that we use, transformer-based model, it's been trained now to understand context.
00:20:57.680 --> 00:21:06.640
Now there is room for more improvement and more training, you know, the more we train this kinds of model, of course, maybe the better and the more accurate they can get to be.
00:21:06.640 --> 00:21:14.000
But I would say that when it comes to understanding context, there's still, you know, some opportunities to do better.
00:21:14.000 --> 00:21:17.680
But I think there are advanced models now that can understand context.
00:21:17.680 --> 00:21:20.640
And of course, it's also based on what we train it with.
00:21:20.640 --> 00:21:24.000
So the more we are able to train it, the better it would it would become.
00:21:24.000 --> 00:21:26.240
But again, there's been a lot of advancement now.
00:21:26.640 --> 00:21:27.920
Outstanding, or fire.
00:21:28.319 --> 00:21:28.559
Yes (laughter).
00:21:31.039 --> 00:21:31.920
Very good.
00:21:31.920 --> 00:21:37.680
Let's, let's turn to the the next study that, that you've had recently accepted.
00:21:37.680 --> 00:21:47.279
So you had this really interesting study that examined mechanisms that identify genetic risk factors for high blood pressure.
00:21:47.279 --> 00:21:53.200
So, similar to my previous question, I'm gonna be honest, you - you exceeded my cognitive capacity in this particular study.
00:21:53.200 --> 00:21:59.839
So can you kind of, could you kind of explain what what you were going for with this, this particular one and what you found?
00:22:00.240 --> 00:22:00.880
Yes.
00:22:00.880 --> 00:22:09.039
You know, like as you mentioned earlier, we have a whole lot of data and try to make sense of it.
00:22:09.039 --> 00:22:09.680
It's ...
00:22:09.680 --> 00:22:11.599
is increasingly important.
00:22:11.599 --> 00:22:17.119
So it's more like to put it in simple terms, we're looking for needle in a haystack.
00:22:17.759 --> 00:22:30.640
We have all of these genetic features that are potential risk factors that are potentially associated with high blood pressure, an ailment that could have detrimental consequences, right?
00:22:30.640 --> 00:22:43.200
And so what we're trying to do in this research is, how do we create a targeted list of features that are associated with high blood pressure?
00:22:43.359 --> 00:22:43.599
Okay.
00:22:43.759 --> 00:22:44.079
Right.
00:22:44.079 --> 00:22:49.440
We have this hundreds of thousands of genetic features.
00:22:49.440 --> 00:23:02.000
If we decide to do analysis with that, well, not only is this going to be, well, with advancement now, I'm not going to say it's going to be inefficient, but it could be some form of waste of time in case you find, you know, several features that are not even relevant.
00:23:02.000 --> 00:23:06.720
And most of the time in studies like this, there's only a small set that is actually relevant.
00:23:06.720 --> 00:23:21.119
So our goal here is can we trim down this such that we now have, you know, a sizable or a targeted list of features that are that have the most likelihood of being relevant to this ailment.
00:23:21.119 --> 00:23:28.400
So that way we can better pursue or dig deeper into this view than having to do with this plethorial features.
00:23:28.640 --> 00:23:28.799
Okay.
00:23:28.960 --> 00:23:30.400
And that's really what we're looking at.
00:23:30.400 --> 00:23:43.039
One of the, you know, amongst other things, why we're motivated to do this is there's been studies that have looked at, hey, what genetic features are associated with blood pressure on an average level.
00:23:43.039 --> 00:23:46.480
We realize that blood pressure is heterogeneous.
00:23:46.480 --> 00:23:54.400
What I mean by that is genetic features that might be relevant to one person may not be for another person.
00:23:54.400 --> 00:24:05.359
And genetic features that might be important for average blood level, blood pressure level may or may not be relevant for higher levels of blood pressure, unusually extreme levels of blood pressure.
00:24:05.519 --> 00:24:05.759
Okay.
00:24:06.000 --> 00:24:10.160
And so we are focusing more on looking at these high levels of blood pressure.
00:24:10.160 --> 00:24:13.279
We believe that they have the most severe consequences.
00:24:13.279 --> 00:24:32.799
And so we developed an efficient algorithm with which we can then efficiently screen all of these features, millions, thousands of them, and kind of narrow them down to those that are most likely associated with high blood pressure, high levels of blood pressure, while limiting the number of false discoveries.
00:24:32.799 --> 00:24:43.440
And what I mean by that is, you know, because we have so much data, there might be some features that are retained as important just by chance or coincidence or due to statistical noise.
00:24:43.440 --> 00:24:50.079
And so by allowing control for false discovery, we're able to limit this as much as possible.
00:24:50.079 --> 00:25:01.279
So at the end of the day, we have a sizable set of features that are associated, potentially associated with blood pressure, and we have limited the number of those that are false discoveries.
00:25:01.599 --> 00:25:01.920
Okay. What, what data go into that, that discovery? I mean how do you get to that point where you can say, these are the ones?
00:25:01.609 --> 00:25:01.920
Yes.
00:25:13.279 --> 00:25:24.799
As the data set that we are using (indistinct) from study data. And we're starting out with, you know, about over 500,000 genetic features.
00:25:24.799 --> 00:25:34.960
And then the goal is, hey, how can we efficiently screen all of these features to get down to those that have plausible, you know, potential association?
00:25:34.960 --> 00:25:45.200
And we put, you know, we put forward a method that not only would look at the average level of blood pressure, but can look at the entire distribution of blood pressure.
00:25:45.200 --> 00:25:53.359
Again, our method can work for any part of this distribution, but we're focusing on high blood pressure, you know, as in the paper.
00:25:53.359 --> 00:26:00.319
And then, you know, like I said earlier, we're limiting for discovery rate.
00:26:00.319 --> 00:26:15.279
And so our outcome, our at the end of the day, the goal is, hey, now we have this genetic, uh, genetic features, and then we get to look into the literature to see, hey, which ones have been identified earlier to be related to blood pressure?
00:26:15.279 --> 00:26:21.680
We're able to find some of them, which then gives us some level of confidence in the performance of our method.
00:26:21.680 --> 00:26:32.160
But at the same time, we're also able to discover newer snips or neurogenetic features that have that don't have prior evidence or little evidence as associated with blood pressure.
00:26:32.160 --> 00:26:37.279
And that's the point where we can then say, hey, we should further investigate these snips.
00:26:37.279 --> 00:26:41.440
Maybe there are potential genetic features that could help us eventually.
00:26:41.440 --> 00:26:49.440
The ultimate goal then is by understanding this, we're able to, you know, develop some more targeted interventions for blood pressure.
00:26:49.440 --> 00:26:56.799
We're also able to look at some strategies with which we can, you know, help, you know, people with high blood pressure.
00:26:56.799 --> 00:26:58.480
At least that's the bigger picture.
00:26:58.720 --> 00:26:58.880
Yeah.
00:26:58.880 --> 00:27:04.240
And it seems like to me that the, the big innovation here is this the method, right?
00:27:04.240 --> 00:27:09.440
So I mean, you applied it to high blood pressure, but you could think of countless different ailments.
00:27:09.759 --> 00:27:10.160
Oh, yes.
00:27:10.480 --> 00:27:14.799
Or, or in some ways non-ailments, because you said it's the entire distribution, right?
00:27:14.799 --> 00:27:19.119
So you could, you could maybe look at positive features or something like that if you were so inclined.
00:27:19.119 --> 00:27:27.359
But that seems to be like the real innovation here is you've, you've come up with a method to sort of efficiently do this to guide future work.
00:27:27.759 --> 00:27:27.920
Right.
00:27:27.920 --> 00:27:33.519
I would say, you know, yes, it's applicable beyond blood pressure, even though that's just our illustrative example.
00:27:33.519 --> 00:27:48.079
But then just think about hey, you're - you have this large data set, you're interested in understanding the patterns in the data, see if there's signals in the data, and you're interested in the specific part of the distribution, right, of your outcome variable.
00:27:48.079 --> 00:27:50.079
This is a method that you can then apply.
00:27:50.079 --> 00:28:03.200
But in line with that, you're also saying, hey, I want to limit the number of features that get into my the number of features that I get at the end of the day in terms of I don't want to see as much coincidences.
00:28:03.200 --> 00:28:11.039
I want to limit the number of noise that get into this final list of features that are relevant, then this is a method that you want to apply.
00:28:11.039 --> 00:28:13.119
So again, it goes beyond just blood pressure.
00:28:13.279 --> 00:28:13.359
Yeah.
00:28:13.519 --> 00:28:22.240
It could, you could apply it to, you know, similar problems where you're looking at filtering down data while still retaining as much signal as possible.
00:28:22.240 --> 00:28:23.839
And of course limiting the noise.
00:28:24.160 --> 00:28:24.400
Right.
00:28:24.400 --> 00:28:25.920
Well, that's really cool.
00:28:25.920 --> 00:28:28.000
Well, let's talk about what's up next for you.
00:28:28.000 --> 00:28:33.039
What are you working on right now that's been interesting and you're, you're looking forward to exploring further?
00:28:33.599 --> 00:28:34.079
Right.
00:28:34.079 --> 00:28:43.839
I'm working on some interesting things, and I think one of it is still in the digital platform space, looking at uncertainty reduction.
00:28:43.839 --> 00:28:49.759
So with a two-sided digital platform, pretty much you have a group of users that you're trying to connect.
00:28:49.759 --> 00:28:58.480
And this, for example, would be Airbnb, Uber, some you know, platforms where they are lenders and borrowers, for example.
00:28:58.480 --> 00:29:03.519
Now, going specifically with the Airbnb, you know, there is some uncertainty involved.
00:29:03.519 --> 00:29:12.720
What that means is as an individual, you're trying to, you know, you're you go on Airbnb, you look at a listing that you're trying to lodge in.
00:29:12.720 --> 00:29:13.759
You have some inf...
00:29:13.759 --> 00:29:18.960
you, there's some information that you see on there, but there's some others that you do not see.
00:29:19.039 --> 00:29:19.200
Yeah.
00:29:19.359 --> 00:29:24.240
In fact, for example, you might not even know the exact spot or location of the unit.
00:29:24.240 --> 00:29:25.839
You might have an idea from the map.
00:29:26.000 --> 00:29:26.240
Right.
00:29:26.400 --> 00:29:26.640
Right?
00:29:26.640 --> 00:29:31.279
But you're you only have access to information that the host has provided for the most part.
00:29:31.279 --> 00:29:33.599
And so there's some uncertainty involved.
00:29:33.599 --> 00:29:43.200
Even though you go ahead and, you know, book this lodging, you do all that online, the reality actually sets in when you actually go offline and actually stay in that unit.
00:29:43.200 --> 00:29:53.519
So something I've been interested in looking at is that, you know, with this uncertainty, by the time you go offline, are there features that the platform has provided?
00:29:53.519 --> 00:30:01.359
Or are there certain traits or things that the host does or provide that helps reduce this level of uncertainty that you feel?
00:30:01.359 --> 00:30:06.319
An example is that, you know, this unit is in a neighborhood.
00:30:06.319 --> 00:30:10.000
By the time you go there in person, you experience the neighborhood.
00:30:10.000 --> 00:30:17.119
How does this help reduce uncertainty that you had before, or maybe not reduce the uncertainty that you had?
00:30:17.119 --> 00:30:21.839
And of course, how does this tie back into the review that you leave online?
00:30:22.319 --> 00:30:25.119
So an element of kind of expectation setting is involved here.
00:30:25.519 --> 00:30:26.720
Yes, that's right.
00:30:26.720 --> 00:30:27.359
Yes.
00:30:27.359 --> 00:30:33.680
And, you know, it's nice that you said that because in line with that, I've also been interested in expectation mismatch.
00:30:33.680 --> 00:30:47.200
You know, besides uncertainty reduction, when you have this information that you glean online on Airbnb platform, host's description of the unit, host's description of the neighborhood, it could set some expectation for the guest.
00:30:47.359 --> 00:30:47.519
Right.
00:30:47.680 --> 00:30:53.200
And so I've been interested in seeing when the guest actually experiences this unit offline.
00:30:53.200 --> 00:30:57.839
Or maybe they're able to experience the hosts offline, depending on the kind of unit that they booked.
00:30:57.839 --> 00:31:00.960
For example, a shared room where you probably shared the unit with the host.
00:31:01.039 --> 00:31:01.359
Oh, okay.
00:31:01.519 --> 00:31:11.759
Or even if you didn't share the unit with the host, you got to see some things about the host, giving you some more idea of who the person might be, you know, what the person likes and things like that.
00:31:11.759 --> 00:31:15.440
Did that help in terms of did it meet your expectation?
00:31:15.440 --> 00:31:19.839
Was it, were you pleasantly surprised, or was it below your expectation?
00:31:19.839 --> 00:31:21.759
So that's something else that I'm interested in.
00:31:22.000 --> 00:31:22.240
Okay.
00:31:22.240 --> 00:31:30.960
And that could help shape maybe recommendations about how, like what you put in the description as a, as a host trying to attract guests.
00:31:31.279 --> 00:31:52.960
Yes, that, that would help because I feel like now not coming from the researcher perspective, but as a person trying to, you know, get into listing or trying to book a unit, I would like, if it were possible, to see some detailed description of what the unit looks like, the neighborhood and things like that.
00:31:52.960 --> 00:31:56.319
So yes, I think that would provide some insights into that.
00:31:56.319 --> 00:32:04.799
And of course, I think by reading that, it could give me some, it might even help, besides expectation, it might even help with my level of uncertainty.
00:32:04.799 --> 00:32:08.640
You know, I feel like I have an idea of what this place is gonna look like.
00:32:08.640 --> 00:32:10.720
I have an idea of what to expect.
00:32:10.880 --> 00:32:11.119
Right.
00:32:11.279 --> 00:32:17.440
But again, we also have to, it's interesting to also look at the other side of things that now I have something to expect.
00:32:17.440 --> 00:32:19.759
Did you really meet what I thought it would be?
00:32:19.920 --> 00:32:20.160
Right.
00:32:20.319 --> 00:32:23.119
Was it below or above, you know, in terms of what I expected?
00:32:23.359 --> 00:32:23.680
Yeah.
00:32:23.680 --> 00:32:28.079
Well, when you crack the code on that, I think there's a consulting business here.
00:32:28.079 --> 00:32:35.440
You can you know, you can - you can sell your services to help better write these, these things to increase bookings.
00:32:35.599 --> 00:32:35.680
Yes.
00:32:35.920 --> 00:32:40.640
And when we - we realize you're, you're no longer working for us, you're just doing that, we'll know what happened.
00:32:40.720 --> 00:32:40.799
Right.
00:32:41.680 --> 00:32:42.079
Very good.
00:32:42.079 --> 00:32:46.319
Well, how about the, the other, the other study with high blood pressure?
00:32:46.319 --> 00:32:50.079
Are there elements of that, that you're kind of taking a look at the next phase on?
00:32:50.480 --> 00:32:51.519
Yes, yes.
00:32:51.519 --> 00:33:12.240
With the prior method, what we had put forward is a method with which we can screen, you know, very complex data with the goal of, you know, retaining a set of genetic features that are potentially associated with blood pressure and reduce the number of noise that is, you know, false discoveries.
00:33:12.240 --> 00:33:19.200
What we are doing now in terms of extending that is looking at a more conditional perspective.
00:33:19.200 --> 00:33:34.400
What I mean by that is we find that with the nature of blood pressure data, and as with many biomedical data sets, is that there might be certain variables that mask this relationship that we're trying to study.
00:33:34.640 --> 00:33:34.960
Okay.
00:33:35.359 --> 00:33:40.880
In an example would be with the NIH freemium at study data that we're working with.
00:33:40.880 --> 00:33:51.039
These data include phenotypes like age of participants, BMI, total cholesterol, and some more information, along with the genetic features that is in the data.
00:33:51.039 --> 00:34:01.680
While our goal is to understand what genetic features are associated with high blood pressure, we also need to understand the interplay and the relationship between these genetic features and the phenotypes.
00:34:01.680 --> 00:34:06.000
And so what we are doing now is accounting for those phenotypes.
00:34:06.000 --> 00:34:14.719
We're controlling for these factors, and we're trying to see if we're able to see a more clearer picture of the genetic features that are actually important.
00:34:14.719 --> 00:34:29.360
In the prior work that we had, we did not necessarily, you know, pay close, you know, in the prior method that we have, we're looking at the relationship between these features and high blood pressure, right?
00:34:29.360 --> 00:34:37.920
But now we're also saying, hey, we have these phenotypes, we have other variables that may have a relationship or, you know, play a role here.
00:34:38.079 --> 00:34:38.239
Right, right.
00:34:38.480 --> 00:34:45.760
So in this case, if we account for them, what does the snip or genetic feature profile look like when we're talking about high blood pressure?
00:34:45.760 --> 00:34:50.320
So that's something else we're working on right now and we're, you know, we're pretty excited about.
00:34:50.559 --> 00:34:50.719
Okay.
00:34:51.039 --> 00:34:53.440
And again, it goes beyond blood pressure.
00:34:53.440 --> 00:35:00.639
This is method, these are methods that could be applied to high complex, you know, high-dimensional complex data.
00:35:00.639 --> 00:35:08.719
As long as the goal is, hey, I have this large data set, I want to filter down to a smaller set that has signal.
00:35:08.719 --> 00:35:11.599
I'm able to retain as much signal in this smaller set.
00:35:11.679 --> 00:35:11.840
Right.
00:35:12.000 --> 00:35:17.599
And I'm able to filter out as much noise or coincidences as possible, as much as possible.
00:35:17.599 --> 00:35:21.679
And again, I'm not just interested in the average level of this outcome variable.
00:35:21.679 --> 00:35:24.880
For example, I'm not just interested in average blood pressure.
00:35:24.880 --> 00:35:28.239
I want to see the different parts of the distribution.
00:35:28.239 --> 00:35:30.800
So this is a method that could be applied as well.
00:35:31.199 --> 00:35:31.679
Very good.
00:35:31.679 --> 00:35:32.320
Very good.
00:35:32.320 --> 00:35:34.719
Well, obviously, your research is keeping you quite busy.
00:35:34.880 --> 00:35:35.280
Oh, yeah.
00:35:36.000 --> 00:35:41.679
But I'm curious, outside of your work for the Carlos Alvarez College of Business, what keeps you busy?
00:35:42.000 --> 00:35:42.639
Oh, yeah.
00:35:42.639 --> 00:35:55.760
While I enjoy the kind of things I study and what I do here at the Alvarez College of Business, when I'm not doing research or I spend time with my family, I just welcomed a baby girl a couple months ago.
00:35:55.760 --> 00:35:58.960
And, she has kept us quite busy.
00:35:59.119 --> 00:35:59.360
Yeah.
00:35:59.599 --> 00:36:03.199
And, you know, it's beautiful to have an addition to our family.
00:36:03.199 --> 00:36:05.199
And so that's what I do.
00:36:05.199 --> 00:36:12.480
And, you know, when I'm not turning to her or we're having some quality family time, you know, I try to get as much rest as I can.
00:36:12.480 --> 00:36:18.320
And of course, you know, if I can steal in a TV show here and there, I would take that as well.
00:36:18.639 --> 00:36:19.519
Yeah, absolutely.
00:36:19.519 --> 00:36:20.800
Very good, very good.
00:36:20.800 --> 00:36:23.840
And as a father three, take the moments when you can get them.
00:36:24.000 --> 00:36:24.159
Right (laughter).
00:36:24.719 --> 00:36:26.079
Yeah, it's especially this time.
00:36:26.079 --> 00:36:29.119
So well, Saidat, it's been an absolute joy talking with you.
00:36:29.119 --> 00:36:42.000
So glad that we have such amazing researchers that are working here in the college, like you, developing new approaches to research that allow us to use the wealth of data that are out there in the world in order to gain some new insights.
00:36:42.000 --> 00:36:44.000
So, thank you for being here today.
00:36:44.000 --> 00:36:53.039
For listeners who might be interested in learning a little bit more about the types of topics we talked about today, and perhaps even pursuing a degree in business analytics.
00:36:53.039 --> 00:36:57.199
As I mentioned, we have a brand new MS in Business Analytics starting this fall.
00:36:57.199 --> 00:37:01.599
Check out business.utsa.edu for more information.
00:37:01.599 --> 00:37:03.920
We'd love to welcome you as a Roadrunner.
00:37:03.920 --> 00:37:07.039
And thank you again, Saidat, for joining me today.
00:37:07.039 --> 00:37:09.920
It's been just such a pleasure, and I really appreciate your time today.
00:37:09.920 --> 00:37:10.239
Thank you.
00:37:10.480 --> 00:37:11.840
Thank you very much for having me.
00:37:11.840 --> 00:37:13.119
Had a great time.
00:37:13.440 --> 00:37:16.480
Thank you for listening to the Inside Alvarez Business Podcast.
00:37:16.480 --> 00:37:23.280
Special thanks to our producer, Brittney Johnson, and for the support of Wendy Frost and Melissa Lackey to help make this podcast possible.
00:37:23.280 --> 00:37:28.400
Stay connected with the Carlos Alvarez College of Business at The University of Texas in San Antonio.
00:37:28.400 --> 00:37:35.840
To learn more about how we are empowering the next generation of business thinkers and conducting groundbreaking research to ensure their success.
00:37:35.840 --> 00:37:42.159
Follow us on social media or visit us online at business.utsa.edu.
00:37:42.159 --> 00:37:45.039
Until next time, I'm Jonathan Halbesleben.
00:37:45.039 --> 00:37:46.000
Thank you for listening.
00:00:00.959 --> 00:00:08.640
You know, the theme behind my work is using data and analytical tools to understand complex problems.
00:00:08.640 --> 00:00:18.719
Try to find patterns in data, understanding why these patterns exist, and what insights that we can glean from this, you know, for practice.
00:00:19.039 --> 00:00:26.000
We live in a data-rich environment these days, but translating data into truly useful information is challenging for business leaders.
00:00:26.000 --> 00:00:29.199
How can we tame data to gain valuable insights?
00:00:29.199 --> 00:00:36.880
We explore this issue and more in this episode of Inside Alvarez Business, a podcast dedicated to bringing you the stories behind the research.
00:00:36.880 --> 00:00:43.840
I'm your host, Jonathon Halbesleben, Dean of the Carlos Alvarez College of Business at The University of Texas at San Antonio.
00:00:43.840 --> 00:00:45.759
Our guest today is Saidat Sanni.
00:00:45.759 --> 00:00:53.920
We had a terrific conversation about how she has developed new approaches to extrapolate information from massive data sets to produce powerful research insights.
00:00:53.920 --> 00:00:57.200
Enjoy this episode of Inside Alvarez Business.
00:00:57.200 --> 00:01:04.879
Today I am joined by Saidat Sanni, assistant professor in the Department of Operations and Analytics in the Carlos Alvarez College of Business.
00:01:04.879 --> 00:01:07.920
She's also an Alvarez Research Fellow here in the college.
00:01:07.920 --> 00:01:15.680
She completed her PhD in business analytics at the University of Cincinnati in 2024 and then joined us shortly after that.
00:01:15.680 --> 00:01:20.000
She's still relatively new in her career, but is already making a pretty huge impact.
00:01:20.000 --> 00:01:26.400
She's recently had papers accepted in some of our leading journals, including Biometrics and Information Systems Research.
00:01:26.400 --> 00:01:34.480
Her work spans disciplinary boundaries, but a common theme is developing machine learning and statistical models for high-dimensional data.
00:01:34.480 --> 00:01:36.640
She's joined us at a perfect time.
00:01:36.640 --> 00:01:40.640
Of course, it would have been a perfect time no matter when you joined us, but you joined us at a perfect time.
00:01:40.640 --> 00:01:46.560
We're launching a new master's in science and business analytics, and so that'll be very helpful to have your expertise as we join that.
00:01:46.560 --> 00:01:49.680
So, Saidat, welcome to Inside Alvarez Business.
00:01:50.000 --> 00:01:51.200
Thank you so much for having me.
00:01:51.200 --> 00:01:53.519
I'm excited to be here and happy to chat.
00:01:53.760 --> 00:01:54.000
Great.
00:01:54.000 --> 00:01:56.560
Well, and I'm looking forward to hearing some of the stories behind your research.
00:01:56.560 --> 00:01:57.840
So this is - this is perfect.
00:01:58.079 --> 00:01:58.560
Yes.
00:01:58.799 --> 00:02:01.120
Well, let's start kind of way at the beginning.
00:02:01.120 --> 00:02:08.879
As frequent listeners to the podcast, which both of them, both frequent listeners, will know.
00:02:08.879 --> 00:02:12.960
I always ask people like why they decided to become a professor.
00:02:12.960 --> 00:02:17.439
Because I don't think there's anybody that grew up saying, I'm going to be a professor of business analytics when I grow up.
00:02:17.439 --> 00:02:20.560
So for you, what - what was the kind of driving factor?
00:02:20.560 --> 00:02:23.520
What was the moment when you decided you wanted to do this for a living?
00:02:23.840 --> 00:02:24.240
Right.
00:02:24.240 --> 00:02:29.520
I think I never, you know, as a kid, I didn't say I wanted to be a professor.
00:02:29.520 --> 00:02:39.439
But I think it just is a combination of things that I naturally enjoy, even though I didn't know earlier on that it was gonna be a career path for me.
00:02:39.439 --> 00:02:46.800
I'm originally from Nigeria, and education is an important part of being Nigerian, you know.
00:02:46.800 --> 00:02:50.719
In my family, it just felt like that's the natural path.
00:02:50.719 --> 00:02:55.680
You know, you get a degree and you know, you just keep going as much as you can.
00:02:55.680 --> 00:03:05.280
And also given that I'm the first child of my family at a young age, I would usually help them with assignments.
00:03:05.280 --> 00:03:07.520
I would tutor my siblings during the holidays.
00:03:07.520 --> 00:03:11.039
So they're busy and not just, you know, they're busy with work.
00:03:11.039 --> 00:03:13.840
You know, my dad would always have me, have me do that.
00:03:13.840 --> 00:03:15.360
And I enjoyed it.
00:03:15.360 --> 00:03:17.439
I went on to high school.
00:03:17.439 --> 00:03:18.960
I continued to do that.
00:03:18.960 --> 00:03:24.479
I enjoyed explaining difficult problems to people, you know, kind of break it down so it's easier.
00:03:24.639 --> 00:03:24.800
Yeah.
00:03:24.960 --> 00:03:30.080
I realized that it also helped me retain the knowledge or the concept.
00:03:30.080 --> 00:03:32.719
Went on to college, continued doing the same thing.
00:03:32.719 --> 00:03:35.680
Even at the time, it wasn't, it just felt like, oh, I enjoyed doing this.
00:03:35.680 --> 00:03:36.719
Oh, can you teach me this?
00:03:36.719 --> 00:03:37.680
Oh, of course I will.
00:03:37.680 --> 00:03:46.240
In terms of then, you know, deciding to be a professor, really started, I started really thinking about that when I started research.
00:03:46.240 --> 00:03:51.039
During my master's program, I'd left Nigeria and moved to the United States for that.
00:03:51.039 --> 00:03:53.039
I started getting into research.
00:03:53.039 --> 00:04:11.199
I enjoyed the idea of having questions and being able to kind of walk through these questions that don't have obvious answers, and then being able to walk through the process of doing the research, solving this problem, and eventually being able to garner some insights from this.
00:04:11.199 --> 00:04:12.560
So I really did like that.
00:04:12.560 --> 00:04:16.959
And then I decided, hey, I'll just get a PhD, you know, might as well.
00:04:16.959 --> 00:04:18.399
I continued research.
00:04:18.399 --> 00:04:22.959
And, you know, in summary, it just kind of fell into place for me.
00:04:22.959 --> 00:04:23.920
I enjoyed teaching.
00:04:23.920 --> 00:04:25.519
Now I enjoy research.
00:04:25.519 --> 00:04:31.040
These two things, going to the career of being a professor, you know, it just felt natural.
00:04:31.600 --> 00:04:32.000
Oh, good.
00:04:32.000 --> 00:04:32.720
Oh, good.
00:04:32.720 --> 00:04:34.240
What a great story.
00:04:34.240 --> 00:04:35.920
So I'm, I'm always curious.
00:04:35.920 --> 00:04:41.600
You so you grew up in Nigeria, and at some point you you made that decision to come to the United States for your master's degree.
00:04:41.759 --> 00:04:41.920
Yes.
00:04:42.160 --> 00:04:43.759
What kind of led to that decision?
00:04:44.079 --> 00:04:44.800
Oh, right.
00:04:44.800 --> 00:04:55.199
Again, it goes back to the culture that my family, or I would say broadly, Nigeria has with always seeking for knowledge.
00:04:55.199 --> 00:05:00.800
You know, after I got the bachelor's degree, I wanted to do something more.
00:05:00.800 --> 00:05:03.680
I felt that, okay, now I have the basics in statistics.
00:05:03.680 --> 00:05:05.759
I have an undergraduate in statistics.
00:05:05.759 --> 00:05:08.240
I can see the practicality of statistics.
00:05:08.240 --> 00:05:10.240
Can I take this a notch farther?
00:05:10.240 --> 00:05:12.639
And then I decided I like the practice.
00:05:12.639 --> 00:05:14.240
How about I get an MBA?
00:05:14.240 --> 00:05:16.000
And so I came to the U.S.
00:05:16.000 --> 00:05:17.519
to get an MBA.
00:05:17.519 --> 00:05:24.560
Really, I just felt, you know, it's the opportunity of being able to advance in that line since I enjoyed this so much.
00:05:24.560 --> 00:05:31.839
Even though, you know, I had people who have supported me along the way, but I also had people who said, this was their opinion, not mine.
00:05:31.839 --> 00:05:34.720
Oh, statistics, it's very technical.
00:05:34.720 --> 00:05:36.399
How do you do business?
00:05:36.399 --> 00:05:41.920
Because then you have to write papers, you have to do presentations and things like that.
00:05:41.920 --> 00:05:44.959
And I just felt, well, that's the practice.
00:05:44.959 --> 00:05:48.319
That's what made me fall in love with statistics in the first place.
00:05:48.319 --> 00:05:51.279
And so that was why I made the decision to move here.
00:05:51.279 --> 00:05:59.360
And after I had the business degree, I decided statistics plus business, let's get a PhD in business analytics.
00:05:59.360 --> 00:06:00.879
And that's how I got there.
00:06:01.120 --> 00:06:01.920
Oh, very good.
00:06:01.920 --> 00:06:02.639
Very good.
00:06:02.639 --> 00:06:04.160
So I, I'm curious then.
00:06:04.160 --> 00:06:13.279
I this might kind of, you sort of answered this question for me, but kind of more specific in terms of the research that you've now pursued since then.
00:06:13.279 --> 00:06:17.360
What, what inspired you to study some of the topics that you've studied?
00:06:17.360 --> 00:06:26.000
We'll talk about a couple of your papers in specific in a little bit here, but just kind of generally speaking, like what got you into some of the areas that you've, you've studied now.
00:06:26.399 --> 00:06:26.959
Right.
00:06:26.959 --> 00:06:38.240
The motivation behind, you know, the theme behind my work is using data and analytical tools to understand complex problems.
00:06:38.240 --> 00:06:48.959
You know, try to find patterns in data, understanding why these patterns exist and what insights that we can glean from this, you know, for practice.
00:06:48.959 --> 00:07:07.360
And majority of my work, while some are methodological, focusing on developing statistical tools to handle complex data, you would also find that I also study technology and human behavior, also driven from you know, looking at complex data sets as well.
00:07:07.360 --> 00:07:09.839
My application then spans in different areas.
00:07:09.839 --> 00:07:27.759
It could be we're going from genetics, which is you know, developing statistical methodology to financial markets, you know, putting forward models to predict stock and bond returns, and of course, digital platforms as well, understanding how technology and interpersonal relationships interplay on online platforms.
00:07:28.160 --> 00:07:36.800
I'm interested in your work in part because I had a mentor kind of early in my career when I was my first job was in a med school, and I didn't know anything about medicine.
00:07:36.800 --> 00:07:38.639
Kind of a problem.
00:07:38.639 --> 00:07:41.439
So I did a lot of work studying in a hospital.
00:07:41.439 --> 00:07:49.519
And one of the things that always stuck with me is I had this mentor who said, This hospital has got tons and tons of data, and we have so little information.
00:07:49.519 --> 00:07:56.959
And so I think you know, folks like you that are kind of taking these huge data sets and trying to distill like what's, what's the real information there?
00:07:56.959 --> 00:07:58.639
I mean, that's, that's - there's a lot of value in that.
00:07:58.639 --> 00:08:01.839
So I'm, I'm excited that we've got you working on some of those things.
00:08:01.839 --> 00:08:06.000
So let's talk about some of your work in a little bit more detail.
00:08:06.000 --> 00:08:11.439
So I'm gonna start with the paper that you've got forthcoming in the journal of Information Systems Research.
00:08:11.439 --> 00:08:12.639
I thought this one was really cool.
00:08:12.639 --> 00:08:33.840
I mean, I thought both papers were talking about were really cool, but ah, so this one it studied how racial similarity between Airbnb hosts and the guests that stay in their properties, of course, how that similarity, the racial similarity impacted the reviews of the guests, which is really kind of an intriguing area.
00:08:33.840 --> 00:08:36.720
I had certainly, I didn't really, hadn't really thought much about that.
00:08:36.720 --> 00:08:56.240
So you had predicted that guests in marginalized pairs, so two like two members of marginalized groups that are together in a pair, they would have more positive reviews, and that then there were some features of the technology that might actually increase that even further, where they have more or more positive reviews, if that's a if that's a phrase.
00:08:56.240 --> 00:09:03.600
So what - I know we were talking about, what inspired you to study these things, but that that's a pretty specific thing to study (laughter).
00:09:03.600 --> 00:09:07.440
So what inspired that kind of really specific thing to take a look at?
00:09:07.759 --> 00:09:08.320
Right.
00:09:08.320 --> 00:09:24.399
Well, digital platforms, you know, be it Airbnb, Uber, or platforms where you have people trying to borrow money, borrowers and lenders, you have, you know, platforms for service providers and you know, people that are trying to use the service, kind of, you know, it's commonplace.
00:09:24.399 --> 00:09:25.600
It's pretty much everywhere.
00:09:25.600 --> 00:09:31.200
And you have people who are then connected on those platforms and they conduct transactions.
00:09:31.200 --> 00:09:46.320
Studies, and even we humans might believe that, you know, if people are similar, you know, they have similar interests, similar lived experiences, similar values, they might likely have some form of familiarity, which could actually, you know, create some form of trust.
00:09:46.320 --> 00:09:54.399
Because to kind of do transactions on those platforms, you need some level of trust to conduct such, such transactions.
00:09:54.639 --> 00:09:56.159
Yeah, I mean you're staying in somebody's home.
00:09:56.480 --> 00:09:56.799
Exactly.
00:09:56.799 --> 00:10:12.720
Like Airbnb, you're staying in somebody's home, someone who you don't have a relationship, you don't know their history, but based on some information that you gleaned from this platform, or maybe if you've had prior conversations with them online on this digital platform, you might, you know, take a chance, you know, doing that.
00:10:12.720 --> 00:10:28.399
And so what really drove the motivation was besides that digital platforms are commonplace, it's to kind of not only say that because people are similar, they have similar interests, then it could lead to perceived positive experience with this transaction or in this digital platform.
00:10:28.399 --> 00:10:30.639
Rather, we're looking at the nuances.
00:10:30.639 --> 00:10:32.639
Maybe similarity is not enough.
00:10:32.639 --> 00:10:44.960
Maybe there's some features that this digital platform has put in place, however simple, however technical, but they have put in place to kind of help this interaction between them.
00:10:44.960 --> 00:10:59.200
And with these technology features, we could see that, you know, similarity could lead to benefits in terms of online, you know, passivity, perceived experience, positive perceived experiences on these platforms.
00:10:59.200 --> 00:11:02.559
But really, what we're looking at is maybe similarity is not enough.
00:11:02.559 --> 00:11:09.279
Maybe there's something else that adds on to that that then allows us to see the benefits of being similar on that platform.
00:11:09.279 --> 00:11:19.600
And so, for example, in Airbnb, you're trying to lodge in somebody's home, you barely know them, but you're conducting this interaction online or this transaction online.
00:11:19.600 --> 00:11:26.879
You go ahead and lodge in their place, something happens probably offline, and then you come back online to kind of give a review.
00:11:26.879 --> 00:11:29.679
The start and end of this relationship happens online.
00:11:29.679 --> 00:11:33.440
So technology is a medium with which these interactions happen.
00:11:33.519 --> 00:11:33.679
Yeah.
00:11:33.919 --> 00:11:46.240
And what we're trying to kind of look at here is how similarity and digital platforms shape perceived experiences on online, pretty much through online reviews.
00:11:46.639 --> 00:11:46.879
Okay.
00:11:46.879 --> 00:11:48.000
And it was interesting.
00:11:48.000 --> 00:11:55.440
So the some of the technology features you looked at were things like the instant booking and like flexible cancellation policies.
00:11:55.440 --> 00:12:05.679
And so it was it was kind of interesting to me that those were features you thought would actually sort of enhance the the effect of the similarity.
00:12:05.679 --> 00:12:11.519
What what led you to that idea that it would be those features that, that would be so important in this?
00:12:12.000 --> 00:12:12.399
Right.
00:12:12.399 --> 00:12:14.559
What we what led to that?
00:12:14.559 --> 00:12:17.039
Well, okay.
00:12:17.039 --> 00:12:23.440
So we found we had, you know, we're thinking, now we're saying going beyond similarity.
00:12:23.440 --> 00:12:33.440
But when you look at similarity at first, even before we go beyond this, we're saying maybe similarities in interests, similarity and lived experiences, similarities and values.
00:12:33.440 --> 00:12:40.399
And we chose this feature specifically because with cancellation, you're looking at flexibility.
00:12:40.559 --> 00:12:40.879
Yeah.
00:12:41.759 --> 00:12:52.159
You have a host who is willing to extend some flexibility to the guest, and a guest who probably likes the idea of, oh, I can cancel at any time, you know, if this arrangement doesn't work for me anymore.
00:12:52.159 --> 00:12:55.519
That's something that it probably may have in common.
00:12:55.679 --> 00:12:55.919
Yeah.
00:12:56.159 --> 00:12:56.559
Right?
00:12:56.559 --> 00:13:00.559
Instant booking, it's we look at it from the lens of shared control.
00:13:00.559 --> 00:13:06.639
What I mean by shared control is that as the host, you have a lot more information about this unit.
00:13:06.639 --> 00:13:08.000
You have more information.
00:13:08.000 --> 00:13:18.720
So if we look at more information as saying you have more control, right, you're also the one who can some of these features of the transaction are controlled by you.
00:13:18.720 --> 00:13:19.759
Who do I accept?
00:13:19.759 --> 00:13:21.120
Who do I not accept?
00:13:21.120 --> 00:13:37.120
By allowing instant booking, this means that with some pre-specified criteria, the guest can actually book to stay in this list in this lodging without having to go through, you know, sending your request and all that.
00:13:37.120 --> 00:13:38.399
It's pretty much instant.
00:13:38.480 --> 00:13:38.639
Right.
00:13:38.799 --> 00:13:42.000
So that way you're sharing some form of control with this person.
00:13:42.000 --> 00:13:46.960
You're allowing that person to have some kind of say-so in this transaction that's going on.
00:13:47.120 --> 00:13:47.279
Okay.
00:13:47.440 --> 00:13:55.840
And so we are chosen this specifically to see that, hey, not only are these people similar in some form, because in terms of similarity, we're looking at social identity here.
00:13:55.840 --> 00:14:12.080
If these two people happen to be similar in that way, if there's now an added feature of, hey, I'm offering some flexibility, I'm offering some shared control, do we see then that the benefits of similarity is weakened or strengthened?
00:14:12.159 --> 00:14:12.320
Yeah.
00:14:12.559 --> 00:14:18.480
And that was really what draw interest in that, you know, something like the host is also giving something that the guest wants.
00:14:18.799 --> 00:14:19.519
Right, right.
00:14:19.519 --> 00:14:32.720
I always I find it interesting that the cancellation policy one I found particularly interesting, mostly because if, if you get to the point where you're leaving a review, you presumably didn't cancel the reservation.
00:14:32.720 --> 00:14:35.840
So it feels a little like almost counterintuitive.
00:14:35.840 --> 00:14:38.399
Like, why would that be involved in this process?
00:14:38.399 --> 00:14:44.960
But I do, I can see that that notion of kind of the shared value of flexibility, and there's some control there too.
00:14:44.960 --> 00:14:46.639
Like I can control if I need to cancel.
00:14:46.639 --> 00:14:49.279
And so that that one was kind of interesting to me.
00:14:49.279 --> 00:14:52.159
Just it felt counterintuitive, but I get it.
00:14:52.159 --> 00:14:53.039
I get it now.
00:14:53.039 --> 00:14:55.519
So thank you for helping me understand.
00:14:55.519 --> 00:15:06.080
So what was interesting, kind of speaking of counterintuitive, you found a little bit of a counterintuitive finding in the sense that you'd expected just the similarity to lead to higher reviews.
00:15:06.080 --> 00:15:15.679
And that wasn't necessarily the case, but once you looked at the features, then the similarity, you know, then there was that effect of higher reviews.
00:15:15.840 --> 00:15:16.000
Right.
00:15:16.240 --> 00:15:17.600
What do you suppose that was?
00:15:18.080 --> 00:15:27.840
I, I guess it just goes back to because people are similar, does not necessarily mean that the interaction is perceived as a positive experience.
00:15:28.080 --> 00:15:28.320
Okay.
00:15:28.559 --> 00:15:33.279
Even though, you know, data platforms try to bring people together, connect them together, which is great.
00:15:33.279 --> 00:15:37.600
But what we are saying here is we should look beyond similarity.
00:15:37.600 --> 00:15:39.440
Similarity is not enough.
00:15:39.600 --> 00:15:39.759
Yeah.
00:15:39.919 --> 00:15:43.200
We need to put forward features.
00:15:43.200 --> 00:15:59.919
Might be simple, but simple features that govern this interaction, such that, you know, it bolsters similarity, but at the same time, it allows people to kind of - the technology aligns with their values, things they appreciate.
00:15:59.919 --> 00:16:04.399
Not to think that, you know, technology is going to have an even effect for everybody.
00:16:04.559 --> 00:16:04.720
Yeah.
00:16:04.879 --> 00:16:07.840
That's really what we're, you know, part of what we're looking at.
00:16:07.840 --> 00:16:11.360
But again, I would say similarity is not enough.
00:16:11.360 --> 00:16:18.240
There are features that govern this transaction such that, you know, both parties kind of enjoy.
00:16:18.240 --> 00:16:24.799
In our case, we're looking at, you know, in our case, we're looking at the guests' experience after the leave a review.
00:16:24.879 --> 00:16:25.200
Right.
00:16:25.360 --> 00:16:43.840
But again, we're saying that this positive reviews that we see are more likely, you know, not just because of similarity, but it's more likely because there's an interplay within, there's an interplay of similarity and the technological features that lead to, you know, positive experiences.
00:16:44.240 --> 00:16:44.559
Okay.
00:16:44.559 --> 00:16:46.320
Interesting.
00:16:46.320 --> 00:16:55.600
One of the things that I think I found really interesting was not just like what you found, but the way you went about finding it.
00:16:55.600 --> 00:17:02.799
So I think one of the things that's maybe hard for a listener to appreciate if they haven't stumbled upon your paper otherwise.
00:17:02.799 --> 00:17:07.200
It's just the massive amount of data you're talking about.
00:17:07.200 --> 00:17:18.000
So you, you were looking at between two cities, New Orleans and New York, over 50,000 listings and like close to 1.5 million reviews.
00:17:18.000 --> 00:17:22.400
And it's not as though you like read through 1.5 million reviews.
00:17:22.400 --> 00:17:25.039
So they like they had a good time.
00:17:25.039 --> 00:17:28.240
So clearly, that's a lot of data to sort through.
00:17:28.240 --> 00:17:31.200
Can you talk a little bit about how you tackled that issue?
00:17:31.200 --> 00:17:33.920
Because I think that's one of the kind of defining features of the study.
00:17:34.240 --> 00:17:38.000
Yes, this was a lot of data, and these were texts.
00:17:38.000 --> 00:17:40.400
These were online reviews that were left by guests.
00:17:40.480 --> 00:17:40.799
Yeah.
00:17:40.960 --> 00:17:47.920
And, you know, yes, the numbers, we had a huge data to work through, huge data to, you know, pre-process.
00:17:47.920 --> 00:17:55.200
Eventually, we in the paper, eventually our findings were based on 500,000 online reviews.
00:17:55.200 --> 00:18:05.039
And what we had done here was apply a transformer-based AI model, distilBERT model, to get the sentiment in these reviews.
00:18:05.039 --> 00:18:13.839
Now we went beyond just counting the number of positive words and negative words to see how positive or how negative this review or each review is.
00:18:13.839 --> 00:18:17.920
What we had done is this transformer-based model is pre-trained.
00:18:17.920 --> 00:18:25.920
And so we're able to see the probability that this is a positive review, neutral, or probability that it's a negative review in terms of sentiment.
00:18:26.079 --> 00:18:26.240
Yeah.
00:18:26.480 --> 00:18:28.880
And, you know, that's really what we applied.
00:18:28.880 --> 00:18:36.960
And then the sentiment then became our outcome measure, our response variable, with which we are now trying to see, hey, what is the effect of similarity on review sentiment?
00:18:36.960 --> 00:18:41.839
What is the effect of, you know, use of this platform features?
00:18:41.839 --> 00:18:53.359
You know, when we, besides similarity, when we look at the interaction of similarity and some of these features, instant booking, a cancellation policy, do we see that reviews are more likely to be positive or otherwise?
00:18:53.359 --> 00:19:05.039
And what we did find is, you know, what we did find from this study is that, again, when we look at the interaction with these features, instant booking, then we find that the reviews are more likely to be positive.
00:19:05.039 --> 00:19:06.720
So it goes beyond just being similar.
00:19:06.720 --> 00:19:24.799
But as an add-on, when you have this policies or not policies, when you have these features that allow for some form of flexibility, that have some form of shared control, that have that aligns more with this us and guest experiences, then we tend to see more positive reviews.
00:19:25.119 --> 00:19:25.359
Yeah.
00:19:25.359 --> 00:19:26.559
So I'm really curious.
00:19:26.559 --> 00:19:31.920
This is gonna demonstrate my, my lack of understanding in some ways.
00:19:31.920 --> 00:19:38.240
So I'm thinking, you know, you've got this machine learning model that's reading through these reviews.
00:19:38.240 --> 00:19:44.960
Presumably it's trained to sort of understand context and almost like slang.
00:19:44.960 --> 00:19:53.200
I mean, I'm thinking when I talk to my kids and I'll say something that's, you know, they'll they'll say something that's really great and they'll be like, yeah, that's fire.
00:19:53.200 --> 00:19:56.480
That would be terrible in an Airbnb review.
00:19:56.720 --> 00:19:57.359
Right?
00:19:58.079 --> 00:20:04.000
You know, or I had a killer time, you know, like that'd be a really problematic in an Airbnb review.
00:20:04.000 --> 00:20:18.160
If, if it the context wasn't kind of there, so the models are they're trained to sort of understand these types of nuances in human communication, or is there, I guess, how do you train the model to understand those nuances?
00:20:18.480 --> 00:20:18.720
Right.
00:20:18.720 --> 00:20:20.400
That that's an excellent question.
00:20:20.400 --> 00:20:25.359
I think that a lot of models have advanced, you know, there've been a lot of advancement.
00:20:25.359 --> 00:20:33.839
Earlier, you know, when you when you're looking at review sentiment, you could count the number of positive words.
00:20:33.839 --> 00:20:37.200
So you could do one word at a time, pretty much dictionary-based.
00:20:37.440 --> 00:20:37.839
Right, right.
00:20:37.839 --> 00:20:38.400
Right.
00:20:38.640 --> 00:20:41.119
And then, you know, that loses the context.
00:20:41.119 --> 00:20:47.759
Because then, you know, when you break up words, you lose the meaning, the the meaning that it gives as an entire sentence.
00:20:47.920 --> 00:20:48.319
Right, right.
00:20:48.480 --> 00:20:48.799
Right?
00:20:48.799 --> 00:20:51.440
But a lot of models have advanced beyond that now.
00:20:51.440 --> 00:20:57.680
Like the model that we use, transformer-based model, it's been trained now to understand context.
00:20:57.680 --> 00:21:06.640
Now there is room for more improvement and more training, you know, the more we train this kinds of model, of course, maybe the better and the more accurate they can get to be.
00:21:06.640 --> 00:21:14.000
But I would say that when it comes to understanding context, there's still, you know, some opportunities to do better.
00:21:14.000 --> 00:21:17.680
But I think there are advanced models now that can understand context.
00:21:17.680 --> 00:21:20.640
And of course, it's also based on what we train it with.
00:21:20.640 --> 00:21:24.000
So the more we are able to train it, the better it would it would become.
00:21:24.000 --> 00:21:26.240
But again, there's been a lot of advancement now.
00:21:26.640 --> 00:21:27.920
Outstanding, or fire.
00:21:28.319 --> 00:21:28.559
Yes (laughter).
00:21:31.039 --> 00:21:31.920
Very good.
00:21:31.920 --> 00:21:37.680
Let's, let's turn to the the next study that, that you've had recently accepted.
00:21:37.680 --> 00:21:47.279
So you had this really interesting study that examined mechanisms that identify genetic risk factors for high blood pressure.
00:21:47.279 --> 00:21:53.200
So, similar to my previous question, I'm gonna be honest, you - you exceeded my cognitive capacity in this particular study.
00:21:53.200 --> 00:21:59.839
So can you kind of, could you kind of explain what what you were going for with this, this particular one and what you found?
00:22:00.240 --> 00:22:00.880
Yes.
00:22:00.880 --> 00:22:09.039
You know, like as you mentioned earlier, we have a whole lot of data and try to make sense of it.
00:22:09.039 --> 00:22:09.680
It's ...
00:22:09.680 --> 00:22:11.599
is increasingly important.
00:22:11.599 --> 00:22:17.119
So it's more like to put it in simple terms, we're looking for needle in a haystack.
00:22:17.759 --> 00:22:30.640
We have all of these genetic features that are potential risk factors that are potentially associated with high blood pressure, an ailment that could have detrimental consequences, right?
00:22:30.640 --> 00:22:43.200
And so what we're trying to do in this research is, how do we create a targeted list of features that are associated with high blood pressure?
00:22:43.359 --> 00:22:43.599
Okay.
00:22:43.759 --> 00:22:44.079
Right.
00:22:44.079 --> 00:22:49.440
We have this hundreds of thousands of genetic features.
00:22:49.440 --> 00:23:02.000
If we decide to do analysis with that, well, not only is this going to be, well, with advancement now, I'm not going to say it's going to be inefficient, but it could be some form of waste of time in case you find, you know, several features that are not even relevant.
00:23:02.000 --> 00:23:06.720
And most of the time in studies like this, there's only a small set that is actually relevant.
00:23:06.720 --> 00:23:21.119
So our goal here is can we trim down this such that we now have, you know, a sizable or a targeted list of features that are that have the most likelihood of being relevant to this ailment.
00:23:21.119 --> 00:23:28.400
So that way we can better pursue or dig deeper into this view than having to do with this plethorial features.
00:23:28.640 --> 00:23:28.799
Okay.
00:23:28.960 --> 00:23:30.400
And that's really what we're looking at.
00:23:30.400 --> 00:23:43.039
One of the, you know, amongst other things, why we're motivated to do this is there's been studies that have looked at, hey, what genetic features are associated with blood pressure on an average level.
00:23:43.039 --> 00:23:46.480
We realize that blood pressure is heterogeneous.
00:23:46.480 --> 00:23:54.400
What I mean by that is genetic features that might be relevant to one person may not be for another person.
00:23:54.400 --> 00:24:05.359
And genetic features that might be important for average blood level, blood pressure level may or may not be relevant for higher levels of blood pressure, unusually extreme levels of blood pressure.
00:24:05.519 --> 00:24:05.759
Okay.
00:24:06.000 --> 00:24:10.160
And so we are focusing more on looking at these high levels of blood pressure.
00:24:10.160 --> 00:24:13.279
We believe that they have the most severe consequences.
00:24:13.279 --> 00:24:32.799
And so we developed an efficient algorithm with which we can then efficiently screen all of these features, millions, thousands of them, and kind of narrow them down to those that are most likely associated with high blood pressure, high levels of blood pressure, while limiting the number of false discoveries.
00:24:32.799 --> 00:24:43.440
And what I mean by that is, you know, because we have so much data, there might be some features that are retained as important just by chance or coincidence or due to statistical noise.
00:24:43.440 --> 00:24:50.079
And so by allowing control for false discovery, we're able to limit this as much as possible.
00:24:50.079 --> 00:25:01.279
So at the end of the day, we have a sizable set of features that are associated, potentially associated with blood pressure, and we have limited the number of those that are false discoveries.
00:25:01.599 --> 00:25:01.920
Okay. What, what data go into that, that discovery? I mean how do you get to that point where you can say, these are the ones?
00:25:01.609 --> 00:25:01.920
Yes.
00:25:13.279 --> 00:25:24.799
As the data set that we are using (indistinct) from study data. And we're starting out with, you know, about over 500,000 genetic features.
00:25:24.799 --> 00:25:34.960
And then the goal is, hey, how can we efficiently screen all of these features to get down to those that have plausible, you know, potential association?
00:25:34.960 --> 00:25:45.200
And we put, you know, we put forward a method that not only would look at the average level of blood pressure, but can look at the entire distribution of blood pressure.
00:25:45.200 --> 00:25:53.359
Again, our method can work for any part of this distribution, but we're focusing on high blood pressure, you know, as in the paper.
00:25:53.359 --> 00:26:00.319
And then, you know, like I said earlier, we're limiting for discovery rate.
00:26:00.319 --> 00:26:15.279
And so our outcome, our at the end of the day, the goal is, hey, now we have this genetic, uh, genetic features, and then we get to look into the literature to see, hey, which ones have been identified earlier to be related to blood pressure?
00:26:15.279 --> 00:26:21.680
We're able to find some of them, which then gives us some level of confidence in the performance of our method.
00:26:21.680 --> 00:26:32.160
But at the same time, we're also able to discover newer snips or neurogenetic features that have that don't have prior evidence or little evidence as associated with blood pressure.
00:26:32.160 --> 00:26:37.279
And that's the point where we can then say, hey, we should further investigate these snips.
00:26:37.279 --> 00:26:41.440
Maybe there are potential genetic features that could help us eventually.
00:26:41.440 --> 00:26:49.440
The ultimate goal then is by understanding this, we're able to, you know, develop some more targeted interventions for blood pressure.
00:26:49.440 --> 00:26:56.799
We're also able to look at some strategies with which we can, you know, help, you know, people with high blood pressure.
00:26:56.799 --> 00:26:58.480
At least that's the bigger picture.
00:26:58.720 --> 00:26:58.880
Yeah.
00:26:58.880 --> 00:27:04.240
And it seems like to me that the, the big innovation here is this the method, right?
00:27:04.240 --> 00:27:09.440
So I mean, you applied it to high blood pressure, but you could think of countless different ailments.
00:27:09.759 --> 00:27:10.160
Oh, yes.
00:27:10.480 --> 00:27:14.799
Or, or in some ways non-ailments, because you said it's the entire distribution, right?
00:27:14.799 --> 00:27:19.119
So you could, you could maybe look at positive features or something like that if you were so inclined.
00:27:19.119 --> 00:27:27.359
But that seems to be like the real innovation here is you've, you've come up with a method to sort of efficiently do this to guide future work.
00:27:27.759 --> 00:27:27.920
Right.
00:27:27.920 --> 00:27:33.519
I would say, you know, yes, it's applicable beyond blood pressure, even though that's just our illustrative example.
00:27:33.519 --> 00:27:48.079
But then just think about hey, you're - you have this large data set, you're interested in understanding the patterns in the data, see if there's signals in the data, and you're interested in the specific part of the distribution, right, of your outcome variable.
00:27:48.079 --> 00:27:50.079
This is a method that you can then apply.
00:27:50.079 --> 00:28:03.200
But in line with that, you're also saying, hey, I want to limit the number of features that get into my the number of features that I get at the end of the day in terms of I don't want to see as much coincidences.
00:28:03.200 --> 00:28:11.039
I want to limit the number of noise that get into this final list of features that are relevant, then this is a method that you want to apply.
00:28:11.039 --> 00:28:13.119
So again, it goes beyond just blood pressure.
00:28:13.279 --> 00:28:13.359
Yeah.
00:28:13.519 --> 00:28:22.240
It could, you could apply it to, you know, similar problems where you're looking at filtering down data while still retaining as much signal as possible.
00:28:22.240 --> 00:28:23.839
And of course limiting the noise.
00:28:24.160 --> 00:28:24.400
Right.
00:28:24.400 --> 00:28:25.920
Well, that's really cool.
00:28:25.920 --> 00:28:28.000
Well, let's talk about what's up next for you.
00:28:28.000 --> 00:28:33.039
What are you working on right now that's been interesting and you're, you're looking forward to exploring further?
00:28:33.599 --> 00:28:34.079
Right.
00:28:34.079 --> 00:28:43.839
I'm working on some interesting things, and I think one of it is still in the digital platform space, looking at uncertainty reduction.
00:28:43.839 --> 00:28:49.759
So with a two-sided digital platform, pretty much you have a group of users that you're trying to connect.
00:28:49.759 --> 00:28:58.480
And this, for example, would be Airbnb, Uber, some you know, platforms where they are lenders and borrowers, for example.
00:28:58.480 --> 00:29:03.519
Now, going specifically with the Airbnb, you know, there is some uncertainty involved.
00:29:03.519 --> 00:29:12.720
What that means is as an individual, you're trying to, you know, you're you go on Airbnb, you look at a listing that you're trying to lodge in.
00:29:12.720 --> 00:29:13.759
You have some inf...
00:29:13.759 --> 00:29:18.960
you, there's some information that you see on there, but there's some others that you do not see.
00:29:19.039 --> 00:29:19.200
Yeah.
00:29:19.359 --> 00:29:24.240
In fact, for example, you might not even know the exact spot or location of the unit.
00:29:24.240 --> 00:29:25.839
You might have an idea from the map.
00:29:26.000 --> 00:29:26.240
Right.
00:29:26.400 --> 00:29:26.640
Right?
00:29:26.640 --> 00:29:31.279
But you're you only have access to information that the host has provided for the most part.
00:29:31.279 --> 00:29:33.599
And so there's some uncertainty involved.
00:29:33.599 --> 00:29:43.200
Even though you go ahead and, you know, book this lodging, you do all that online, the reality actually sets in when you actually go offline and actually stay in that unit.
00:29:43.200 --> 00:29:53.519
So something I've been interested in looking at is that, you know, with this uncertainty, by the time you go offline, are there features that the platform has provided?
00:29:53.519 --> 00:30:01.359
Or are there certain traits or things that the host does or provide that helps reduce this level of uncertainty that you feel?
00:30:01.359 --> 00:30:06.319
An example is that, you know, this unit is in a neighborhood.
00:30:06.319 --> 00:30:10.000
By the time you go there in person, you experience the neighborhood.
00:30:10.000 --> 00:30:17.119
How does this help reduce uncertainty that you had before, or maybe not reduce the uncertainty that you had?
00:30:17.119 --> 00:30:21.839
And of course, how does this tie back into the review that you leave online?
00:30:22.319 --> 00:30:25.119
So an element of kind of expectation setting is involved here.
00:30:25.519 --> 00:30:26.720
Yes, that's right.
00:30:26.720 --> 00:30:27.359
Yes.
00:30:27.359 --> 00:30:33.680
And, you know, it's nice that you said that because in line with that, I've also been interested in expectation mismatch.
00:30:33.680 --> 00:30:47.200
You know, besides uncertainty reduction, when you have this information that you glean online on Airbnb platform, host's description of the unit, host's description of the neighborhood, it could set some expectation for the guest.
00:30:47.359 --> 00:30:47.519
Right.
00:30:47.680 --> 00:30:53.200
And so I've been interested in seeing when the guest actually experiences this unit offline.
00:30:53.200 --> 00:30:57.839
Or maybe they're able to experience the hosts offline, depending on the kind of unit that they booked.
00:30:57.839 --> 00:31:00.960
For example, a shared room where you probably shared the unit with the host.
00:31:01.039 --> 00:31:01.359
Oh, okay.
00:31:01.519 --> 00:31:11.759
Or even if you didn't share the unit with the host, you got to see some things about the host, giving you some more idea of who the person might be, you know, what the person likes and things like that.
00:31:11.759 --> 00:31:15.440
Did that help in terms of did it meet your expectation?
00:31:15.440 --> 00:31:19.839
Was it, were you pleasantly surprised, or was it below your expectation?
00:31:19.839 --> 00:31:21.759
So that's something else that I'm interested in.
00:31:22.000 --> 00:31:22.240
Okay.
00:31:22.240 --> 00:31:30.960
And that could help shape maybe recommendations about how, like what you put in the description as a, as a host trying to attract guests.
00:31:31.279 --> 00:31:52.960
Yes, that, that would help because I feel like now not coming from the researcher perspective, but as a person trying to, you know, get into listing or trying to book a unit, I would like, if it were possible, to see some detailed description of what the unit looks like, the neighborhood and things like that.
00:31:52.960 --> 00:31:56.319
So yes, I think that would provide some insights into that.
00:31:56.319 --> 00:32:04.799
And of course, I think by reading that, it could give me some, it might even help, besides expectation, it might even help with my level of uncertainty.
00:32:04.799 --> 00:32:08.640
You know, I feel like I have an idea of what this place is gonna look like.
00:32:08.640 --> 00:32:10.720
I have an idea of what to expect.
00:32:10.880 --> 00:32:11.119
Right.
00:32:11.279 --> 00:32:17.440
But again, we also have to, it's interesting to also look at the other side of things that now I have something to expect.
00:32:17.440 --> 00:32:19.759
Did you really meet what I thought it would be?
00:32:19.920 --> 00:32:20.160
Right.
00:32:20.319 --> 00:32:23.119
Was it below or above, you know, in terms of what I expected?
00:32:23.359 --> 00:32:23.680
Yeah.
00:32:23.680 --> 00:32:28.079
Well, when you crack the code on that, I think there's a consulting business here.
00:32:28.079 --> 00:32:35.440
You can you know, you can - you can sell your services to help better write these, these things to increase bookings.
00:32:35.599 --> 00:32:35.680
Yes.
00:32:35.920 --> 00:32:40.640
And when we - we realize you're, you're no longer working for us, you're just doing that, we'll know what happened.
00:32:40.720 --> 00:32:40.799
Right.
00:32:41.680 --> 00:32:42.079
Very good.
00:32:42.079 --> 00:32:46.319
Well, how about the, the other, the other study with high blood pressure?
00:32:46.319 --> 00:32:50.079
Are there elements of that, that you're kind of taking a look at the next phase on?
00:32:50.480 --> 00:32:51.519
Yes, yes.
00:32:51.519 --> 00:33:12.240
With the prior method, what we had put forward is a method with which we can screen, you know, very complex data with the goal of, you know, retaining a set of genetic features that are potentially associated with blood pressure and reduce the number of noise that is, you know, false discoveries.
00:33:12.240 --> 00:33:19.200
What we are doing now in terms of extending that is looking at a more conditional perspective.
00:33:19.200 --> 00:33:34.400
What I mean by that is we find that with the nature of blood pressure data, and as with many biomedical data sets, is that there might be certain variables that mask this relationship that we're trying to study.
00:33:34.640 --> 00:33:34.960
Okay.
00:33:35.359 --> 00:33:40.880
In an example would be with the NIH freemium at study data that we're working with.
00:33:40.880 --> 00:33:51.039
These data include phenotypes like age of participants, BMI, total cholesterol, and some more information, along with the genetic features that is in the data.
00:33:51.039 --> 00:34:01.680
While our goal is to understand what genetic features are associated with high blood pressure, we also need to understand the interplay and the relationship between these genetic features and the phenotypes.
00:34:01.680 --> 00:34:06.000
And so what we are doing now is accounting for those phenotypes.
00:34:06.000 --> 00:34:14.719
We're controlling for these factors, and we're trying to see if we're able to see a more clearer picture of the genetic features that are actually important.
00:34:14.719 --> 00:34:29.360
In the prior work that we had, we did not necessarily, you know, pay close, you know, in the prior method that we have, we're looking at the relationship between these features and high blood pressure, right?
00:34:29.360 --> 00:34:37.920
But now we're also saying, hey, we have these phenotypes, we have other variables that may have a relationship or, you know, play a role here.
00:34:38.079 --> 00:34:38.239
Right, right.
00:34:38.480 --> 00:34:45.760
So in this case, if we account for them, what does the snip or genetic feature profile look like when we're talking about high blood pressure?
00:34:45.760 --> 00:34:50.320
So that's something else we're working on right now and we're, you know, we're pretty excited about.
00:34:50.559 --> 00:34:50.719
Okay.
00:34:51.039 --> 00:34:53.440
And again, it goes beyond blood pressure.
00:34:53.440 --> 00:35:00.639
This is method, these are methods that could be applied to high complex, you know, high-dimensional complex data.
00:35:00.639 --> 00:35:08.719
As long as the goal is, hey, I have this large data set, I want to filter down to a smaller set that has signal.
00:35:08.719 --> 00:35:11.599
I'm able to retain as much signal in this smaller set.
00:35:11.679 --> 00:35:11.840
Right.
00:35:12.000 --> 00:35:17.599
And I'm able to filter out as much noise or coincidences as possible, as much as possible.
00:35:17.599 --> 00:35:21.679
And again, I'm not just interested in the average level of this outcome variable.
00:35:21.679 --> 00:35:24.880
For example, I'm not just interested in average blood pressure.
00:35:24.880 --> 00:35:28.239
I want to see the different parts of the distribution.
00:35:28.239 --> 00:35:30.800
So this is a method that could be applied as well.
00:35:31.199 --> 00:35:31.679
Very good.
00:35:31.679 --> 00:35:32.320
Very good.
00:35:32.320 --> 00:35:34.719
Well, obviously, your research is keeping you quite busy.
00:35:34.880 --> 00:35:35.280
Oh, yeah.
00:35:36.000 --> 00:35:41.679
But I'm curious, outside of your work for the Carlos Alvarez College of Business, what keeps you busy?
00:35:42.000 --> 00:35:42.639
Oh, yeah.
00:35:42.639 --> 00:35:55.760
While I enjoy the kind of things I study and what I do here at the Alvarez College of Business, when I'm not doing research or I spend time with my family, I just welcomed a baby girl a couple months ago.
00:35:55.760 --> 00:35:58.960
And, she has kept us quite busy.
00:35:59.119 --> 00:35:59.360
Yeah.
00:35:59.599 --> 00:36:03.199
And, you know, it's beautiful to have an addition to our family.
00:36:03.199 --> 00:36:05.199
And so that's what I do.
00:36:05.199 --> 00:36:12.480
And, you know, when I'm not turning to her or we're having some quality family time, you know, I try to get as much rest as I can.
00:36:12.480 --> 00:36:18.320
And of course, you know, if I can steal in a TV show here and there, I would take that as well.
00:36:18.639 --> 00:36:19.519
Yeah, absolutely.
00:36:19.519 --> 00:36:20.800
Very good, very good.
00:36:20.800 --> 00:36:23.840
And as a father three, take the moments when you can get them.
00:36:24.000 --> 00:36:24.159
Right (laughter).
00:36:24.719 --> 00:36:26.079
Yeah, it's especially this time.
00:36:26.079 --> 00:36:29.119
So well, Saidat, it's been an absolute joy talking with you.
00:36:29.119 --> 00:36:42.000
So glad that we have such amazing researchers that are working here in the college, like you, developing new approaches to research that allow us to use the wealth of data that are out there in the world in order to gain some new insights.
00:36:42.000 --> 00:36:44.000
So, thank you for being here today.
00:36:44.000 --> 00:36:53.039
For listeners who might be interested in learning a little bit more about the types of topics we talked about today, and perhaps even pursuing a degree in business analytics.
00:36:53.039 --> 00:36:57.199
As I mentioned, we have a brand new MS in Business Analytics starting this fall.
00:36:57.199 --> 00:37:01.599
Check out business.utsa.edu for more information.
00:37:01.599 --> 00:37:03.920
We'd love to welcome you as a Roadrunner.
00:37:03.920 --> 00:37:07.039
And thank you again, Saidat, for joining me today.
00:37:07.039 --> 00:37:09.920
It's been just such a pleasure, and I really appreciate your time today.
00:37:09.920 --> 00:37:10.239
Thank you.
00:37:10.480 --> 00:37:11.840
Thank you very much for having me.
00:37:11.840 --> 00:37:13.119
Had a great time.
00:37:13.440 --> 00:37:16.480
Thank you for listening to the Inside Alvarez Business Podcast.
00:37:16.480 --> 00:37:23.280
Special thanks to our producer, Brittney Johnson, and for the support of Wendy Frost and Melissa Lackey to help make this podcast possible.
00:37:23.280 --> 00:37:28.400
Stay connected with the Carlos Alvarez College of Business at The University of Texas in San Antonio.
00:37:28.400 --> 00:37:35.840
To learn more about how we are empowering the next generation of business thinkers and conducting groundbreaking research to ensure their success.
00:37:35.840 --> 00:37:42.159
Follow us on social media or visit us online at business.utsa.edu.
00:37:42.159 --> 00:37:45.039
Until next time, I'm Jonathan Halbesleben.
00:37:45.039 --> 00:37:46.000
Thank you for listening.
