Network-based learning for understanding collective human behaviour

1
00:00:08,810 --> 00:00:15,980
Today, we're excited to hear from Guangdong on network based learning for understanding collective human behaviour.

2
00:00:15,980 --> 00:00:26,990
He's a lecturer in the Department of Engineering here at Oxford and a former postdoc in The Last Media Lab, where he is still an affiliate shoutin,

3
00:00:26,990 --> 00:00:37,760
works with graphs to model relational structure of data and develop new technologies that sit at the intersection of machine learning,

4
00:00:37,760 --> 00:00:41,540
signal processing and complex networks. So it's kind of a cool opportunity.

5
00:00:41,540 --> 00:00:50,450
We've touched quite a bit peripherally on mentioning networks, but now we're going to get to hear a little bit more in depth.

6
00:00:50,450 --> 00:01:01,430
And we're also lucky to be creating. Speaking of networks, a little bit of traffic closer here because at MIT Media Lab, he was in Sandy Petaling lab,

7
00:01:01,430 --> 00:01:11,300
which several friends and former members of of the six informal in other years and speakers this year, like Abdullah speaking at Princeton.

8
00:01:11,300 --> 00:01:15,180
Mm hmm. So we were creating friends of a friend of a friend.

9
00:01:15,180 --> 00:01:23,220
So, yeah, thank you very much for that saying. Yeah, thank you.

10
00:01:23,220 --> 00:01:30,930
So I hope this this is OK, you can hear it. Thank you very much for the kind introduction and the invitation.

11
00:01:30,930 --> 00:01:36,630
It's a great pleasure to be here and also to talk about some recent stuff that we did.

12
00:01:36,630 --> 00:01:41,400
So I'm in engineering here technically so, which means I have an engineering background.

13
00:01:41,400 --> 00:01:48,330
So that means when I talk about computational social science, I definitely have more on the computational side rather than the social side.

14
00:01:48,330 --> 00:01:57,270
So I will not pretend to be a social scientist here. So if I make an imprecise statement, please forgive me.

15
00:01:57,270 --> 00:02:06,390
Nevertheless, I think it's good to present some recent work that we have done to apply computational approaches to understand

16
00:02:06,390 --> 00:02:15,300
collective human behaviour and which I hope will bring some fresh ideas from a more computational learning perspective,

17
00:02:15,300 --> 00:02:23,850
which would be complementary to the traditional approaches. So for this reason, please feel free to interrupt during the presentation.

18
00:02:23,850 --> 00:02:29,700
If you have any questions or comments, I'll be happy to answer.

19
00:02:29,700 --> 00:02:36,420
So, yes, this works. So since we are in this summer Institute for Computational Social Science,

20
00:02:36,420 --> 00:02:41,130
I thought it would be good for us to first to look at a few historical development

21
00:02:41,130 --> 00:02:45,540
in applying computational methods to understand the society and behaviour.

22
00:02:45,540 --> 00:02:51,390
And you will see most of the many ideas were actually originated from the field of physics.

23
00:02:51,390 --> 00:02:58,590
So in early 19th century and inspired by a Newtonian physics society,

24
00:02:58,590 --> 00:03:08,520
was conceptualised as a machine and Auguste Collins was regarded as the father of sociology by many people at that

25
00:03:08,520 --> 00:03:17,020
time defined the concept of social physics as the study of the laws of societies and the science of the civilisation.

26
00:03:17,020 --> 00:03:27,050
In fact, the term sociology was invented, by our count, to designate the concept of social physics.

27
00:03:27,050 --> 00:03:35,760
And so if we want to study the laws of society, then we need information and, uh, what we call data today.

28
00:03:35,760 --> 00:03:45,810
So at the same time, and to be precise in 1830 for the London Statistical Society Statistical Society of London was founded,

29
00:03:45,810 --> 00:03:48,090
which eventually became the Royal Statistical Society.

30
00:03:48,090 --> 00:03:58,050
Nowadays, with the aim to procure a wrench and publish facts to illustrate the conditions and the prospects of society.

31
00:03:58,050 --> 00:04:02,910
So that was the of data collection. But in Victorian times,

32
00:04:02,910 --> 00:04:07,980
the way that we collected data was basically to go to different villages in this country

33
00:04:07,980 --> 00:04:13,860
and then to collect somewhat disconnected and fragmented pieces of information.

34
00:04:13,860 --> 00:04:19,800
So they're not good enough for us to illustrate the whole picture of the society,

35
00:04:19,800 --> 00:04:25,710
let alone to build computational models to understand that the behaviour of people.

36
00:04:25,710 --> 00:04:28,320
So this is a kind of the first wave of studies.

37
00:04:28,320 --> 00:04:37,740
The second wave took place in the mid-20th century, and again, the concept of social physics was reinvented.

38
00:04:37,740 --> 00:04:46,680
There was actually a research group called Department of Social Physics in Princeton in the 1950s.

39
00:04:46,680 --> 00:04:49,440
They existed for 20 years or something.

40
00:04:49,440 --> 00:04:57,480
So the main drivers of this wave was that many social indicators have been found to have statistical regularities,

41
00:04:57,480 --> 00:05:01,350
such as the zip distribution or gravity law.

42
00:05:01,350 --> 00:05:10,500
So zip distribution is one type of power or distribution, which was popularised by American linguist George Ziff,

43
00:05:10,500 --> 00:05:15,060
who studied the distribution of the frequencies of the words in a document.

44
00:05:15,060 --> 00:05:24,270
And he found that out that the frequencies of a word in a document is proportional to the rank of the word in the frequency table.

45
00:05:24,270 --> 00:05:32,670
So that means the most frequent. The word appears two times as many as often as the second most frequent and three times as many

46
00:05:32,670 --> 00:05:39,780
as the search for frequency and the same distribution has been discovered for other indicators,

47
00:05:39,780 --> 00:05:46,020
such as the populations of cities in countries or income distribution.

48
00:05:46,020 --> 00:05:56,460
And the gravity law is another example of which was also kind of proposed by Ziff, who studied the intercity movement in the United States.

49
00:05:56,460 --> 00:06:01,890
So he he discovered that the number of people who move between two different communities

50
00:06:01,890 --> 00:06:07,020
in the United States is proportional to the population of the two communities.

51
00:06:07,020 --> 00:06:18,880
But inversely proportional to the transportation distance between them. So like a P one, p two divided by D has the name Gravity Law, so.

52
00:06:18,880 --> 00:06:29,380
And then more recently, the third wave emerges where more people find the statistical regularities also within human movement and communication.

53
00:06:29,380 --> 00:06:35,560
So examples could be concepts such as tragic culture.

54
00:06:35,560 --> 00:06:39,400
So how people form a group so like a join groups of form ties.

55
00:06:39,400 --> 00:06:45,010
If a is a friend of both B and C, and B and C are likely to be a friend in future.

56
00:06:45,010 --> 00:06:53,290
Other concepts could be strong and weak ties. So these were the ideas originally proposed by Grant, the latter in the 70s.

57
00:06:53,290 --> 00:06:58,810
So he was basically saying that the links of a president in a social network can be divided in roughly two groups.

58
00:06:58,810 --> 00:07:07,870
The strong ties, which are close friends and weak ties, which are distant acquaintances of the weak ties in this case essentially are

59
00:07:07,870 --> 00:07:13,840
like links between that span the boundaries of different communities or groups,

60
00:07:13,840 --> 00:07:19,570
which are believed to be essential to a exchange of information and creation of new economic opportunities.

61
00:07:19,570 --> 00:07:26,510
So as a consequence, people who maintain what is weak ties are likely to be more successful in their career.

62
00:07:26,510 --> 00:07:33,010
The same idea has been tested at the community level. In a study in 2010 by Ego,

63
00:07:33,010 --> 00:07:38,830
so they basically looked at the communication network of communities across the whole country

64
00:07:38,830 --> 00:07:46,270
of U.K. and then they find the very strong correlation between the diversity of the community,

65
00:07:46,270 --> 00:07:50,940
of the communication patterns of the communities and the socio economic status.

66
00:07:50,940 --> 00:07:53,980
And so this paper itself has also some criticism,

67
00:07:53,980 --> 00:08:00,630
but I think it's one of the first studies that try to test these ideas at the national level, which is good.

68
00:08:00,630 --> 00:08:08,350
But the limitations of this kind of work would be mostly we don't really study the mechanism

69
00:08:08,350 --> 00:08:14,440
behind the social interactions or what is the cause of the observed statistical regularities,

70
00:08:14,440 --> 00:08:19,720
which could be due to a lack of appropriate data or computational methods to do so.

71
00:08:19,720 --> 00:08:28,750
So both things change dramatically more recently. So since 10 years ago, we haven't won more data about human behaviour.

72
00:08:28,750 --> 00:08:38,740
So you can think about social media location data that are that are collected by pervasive technologies such as mobile phones and credit cards.

73
00:08:38,740 --> 00:08:44,200
So one thing about these type of data is that they're passively collected so they are not, like obtained the survey.

74
00:08:44,200 --> 00:08:51,460
So potentially they contain less personal bias because usually the data are not collected

75
00:08:51,460 --> 00:08:56,590
where the people actually are knowing that they're actively being correctly collected.

76
00:08:56,590 --> 00:09:02,170
And also, of course, they are easy to scale to the population scale.

77
00:09:02,170 --> 00:09:07,270
So at the same time, we obviously have more advanced computational methods.

78
00:09:07,270 --> 00:09:17,860
You probably heard about machine learning or deep learning these kind of techniques that can allow us to make the best out of the data that we have.

79
00:09:17,860 --> 00:09:23,770
So the combination of these two basically leads to computational social science,

80
00:09:23,770 --> 00:09:29,590
in my view, which gives a few new perspectives in social science studies.

81
00:09:29,590 --> 00:09:39,250
So we can obviously think about moving away from static information collected by surveys to the actual behaviour of the people.

82
00:09:39,250 --> 00:09:45,690
So from demographics, age, gender to the actual behaviour of the people.

83
00:09:45,690 --> 00:09:49,350
And also from a single prison to a connected network,

84
00:09:49,350 --> 00:09:55,410
so you're not just looking at individuals separately, but you look at how they behave collectively,

85
00:09:55,410 --> 00:10:01,440
you look at this relationship between different people in a dynamic way, which means that the relationship can evolve over time.

86
00:10:01,440 --> 00:10:05,010
And obviously, you can easily look at a much larger population.

87
00:10:05,010 --> 00:10:12,360
So the practical impact of that could be if we think about most current population management solutions,

88
00:10:12,360 --> 00:10:16,200
they're mostly focussed on demographics, individual records and static information.

89
00:10:16,200 --> 00:10:24,450
And think about how banks give credit limits to to you or the network providers to keep videos about mobile phone contracts.

90
00:10:24,450 --> 00:10:32,580
It's mostly about your individual records in history without looking at a kind of a network of people.

91
00:10:32,580 --> 00:10:39,900
So the new way could be to actually look at the actual behaviour of people in a collective way and also dynamically.

92
00:10:39,900 --> 00:10:47,940
But these, of course, need more and more those and the matters because we need to understand better the structures and the complexities.

93
00:10:47,940 --> 00:10:55,320
And and this also poses challenges in developing advanced computational methods in the context of social science research.

94
00:10:55,320 --> 00:10:59,820
So the topic today will mostly be this part.

95
00:10:59,820 --> 00:11:06,840
So how we move from individual behaviour to collective behaviour because I think that will be something important to understand,

96
00:11:06,840 --> 00:11:16,620
also interesting to understand. And so our focus on two pieces of studies, I was pleased to talk equally more or less.

97
00:11:16,620 --> 00:11:20,670
The first question is how social network will affect the decision making.

98
00:11:20,670 --> 00:11:29,280
So we all know that, for example, if I tell you something about a restaurant, maybe you are likely to go sometime in the future.

99
00:11:29,280 --> 00:11:37,020
So that's like an influence. But how this influence propagates in social network is an interesting and open question.

100
00:11:37,020 --> 00:11:44,580
Second, in many cases, we don't observe a social network. We just observe the decisions that are taken by people.

101
00:11:44,580 --> 00:11:49,200
So it will be very interesting to see if we can do the some of the reverse problem.

102
00:11:49,200 --> 00:11:57,030
We observed the actions and then we tried to infer the connexions between them so that we can understand that the communities, et cetera.

103
00:11:57,030 --> 00:12:08,670
So our first talk about that, the the first question. So the kind of motivation here is it's basically we all know that nowadays we live

104
00:12:08,670 --> 00:12:15,180
in a connected society and we want people who are very worried if we are connected.

105
00:12:15,180 --> 00:12:26,130
While few hopes of acquaintances and ideas of the six degrees of separation was mostly first realised by Milgram's experiment in the 70s.

106
00:12:26,130 --> 00:12:35,570
Once you did this, just to ask a present in who lives in salary in the US, Nebraska to to send a letter to somebody in Massachusetts.

107
00:12:35,570 --> 00:12:40,780
And the rule is just to send a letter to one of the people the sender knows.

108
00:12:40,780 --> 00:12:45,570
And so eventually he did a study and then eventually, I think 25 percent of the letters arrived.

109
00:12:45,570 --> 00:12:51,900
But amongst the letters that arrived, the average hope that took is like five point five.

110
00:12:51,900 --> 00:12:57,270
So that was one of the original experiments that gave the idea of six degrees of separation.

111
00:12:57,270 --> 00:13:02,910
And but we are not just connected in a society in like a factual way.

112
00:13:02,910 --> 00:13:05,910
Actually, we also influence each other if we are connected.

113
00:13:05,910 --> 00:13:13,560
So there is several studies by Christakis and Fowler that look at the spreading of behaviours,

114
00:13:13,560 --> 00:13:18,990
certain type of phenomenon in the social network, such as obesity, happiness or smoking behaviour.

115
00:13:18,990 --> 00:13:24,600
And they find that this kind of behaviour can spread two to three hops away from you.

116
00:13:24,600 --> 00:13:28,620
So it's remarkable the way if you think about it may be a person that you don't even know.

117
00:13:28,620 --> 00:13:38,880
In some cases. So the thing that remains open is how this influence propagate in the social network a scenario.

118
00:13:38,880 --> 00:13:45,600
So how might decision will be affected by you if I received some information from you?

119
00:13:45,600 --> 00:13:53,670
And especially in the offline setting, there are a few studies on in the online setting like Facebook, these kind of platforms.

120
00:13:53,670 --> 00:14:01,470
But offline settings are difficult because it's not easy, usually to carry out a large scale offline experiment.

121
00:14:01,470 --> 00:14:09,810
And we also need more advanced methods compared to the traditional asset based methods to to basically quantify the influence.

122
00:14:09,810 --> 00:14:13,380
So that's basically the motivation of this study.

123
00:14:13,380 --> 00:14:26,220
We we looked at a particular dataset, which is a mobile phone dataset collected during an international event called Cirque du Soleil.

124
00:14:26,220 --> 00:14:40,440
It's basically a performance event in many countries. This one is in Andorra in Europe, took place every July, I think, and last for the whole month.

125
00:14:40,440 --> 00:14:45,280
So basically, we try to ask the question.

126
00:14:45,280 --> 00:14:47,110
How influence,

127
00:14:47,110 --> 00:14:56,770
how social influence can play a role in an individual decision making in the sense that whether or not to attend this event is likely to

128
00:14:56,770 --> 00:15:06,580
be affected by the fact that either you receive a phone call from an attendee or maybe somebody who has been contacted by the attending.

129
00:15:06,580 --> 00:15:16,960
These are just some basic statistics that I will explain further to, citing that we have a quite intuitive so we build the information cascade.

130
00:15:16,960 --> 00:15:21,760
So it's basically a phone communication network for a given observational period.

131
00:15:21,760 --> 00:15:29,450
You can take one day if you like. So the idea here is that let's say we have and who is who is the initial adopter.

132
00:15:29,450 --> 00:15:38,980
So he attended the event. And then after that, within this observational period, we observed that she made a phone call to three other people.

133
00:15:38,980 --> 00:15:47,680
OK, then you and Bob, since these people, after receiving the call from and there were further make calls to the other people.

134
00:15:47,680 --> 00:15:51,370
So this is basically the information cascade that we define.

135
00:15:51,370 --> 00:16:01,150
And here we basically define the shortest distance between any person to the original adopter as a hope index.

136
00:16:01,150 --> 00:16:08,650
So you can see Bob Casey and Daniel Day one hop away and all the other people to hopefully the

137
00:16:08,650 --> 00:16:14,170
people who have never been connected in this information cascade are the people who are outside.

138
00:16:14,170 --> 00:16:21,880
So in that sense, we do find all these people to be influenced and the people outside are not influenced, basically.

139
00:16:21,880 --> 00:16:35,260
So then we can observe whether all the rest of the people will make our adoption in this case, attend the event for the rest of the month, basically.

140
00:16:35,260 --> 00:16:39,160
So this is the basic setup.

141
00:16:39,160 --> 00:16:47,800
But one technical challenge in this case would be the decisions on adoption, or a lot of things may be attributed to two factors usually.

142
00:16:47,800 --> 00:16:55,130
One is called Homophily. The Chinese influence the Homophily is basically it means that a.

143
00:16:55,130 --> 00:17:02,570
Friendship is more likely to be formed between people who share certain characteristics in this case or to people like music,

144
00:17:02,570 --> 00:17:10,310
then they're likely to form a link between them instead of between a person who likes music and the person who likes books.

145
00:17:10,310 --> 00:17:20,570
That's the basic idea how links can be formed. So that means if we observe that these two people they both attend, let's say this is attendee.

146
00:17:20,570 --> 00:17:28,130
And then after the phone call, this person goes that we are not really sure whether that person goes because of the phone call,

147
00:17:28,130 --> 00:17:33,080
because it might be that they are just both interested in music by default.

148
00:17:33,080 --> 00:17:38,510
So it's kind of a bias that is caused by this concept of Homophily.

149
00:17:38,510 --> 00:17:47,840
So if we really want to quantify the influence of a phone call and which is the kind of treatment in our case, then we need to separate them.

150
00:17:47,840 --> 00:17:55,970
So this is the other scenario that we would like to consider if all of them are kind of matched based on their interest.

151
00:17:55,970 --> 00:18:00,260
So by default, they are kind of having the same amount of interest of attending the event.

152
00:18:00,260 --> 00:18:09,530
And then in that case, the phone call or basically play the role of affecting the decision making and the most of the studies this this influence,

153
00:18:09,530 --> 00:18:14,450
what we call is actually the combination of two things first, exposure. So I tell you the event.

154
00:18:14,450 --> 00:18:22,070
Second, my experience will basically affects what you do in the future.

155
00:18:22,070 --> 00:18:27,530
Usually, in most of the studies, we don't separate exposure and under the actual influence.

156
00:18:27,530 --> 00:18:40,890
So that's also the case here. OK, so to do this, basically our idea is to kind of use an idea similar to a randomised experiment.

157
00:18:40,890 --> 00:18:45,740
And so we do a matched sample estimation where propensity score matching sounds.

158
00:18:45,740 --> 00:18:55,190
You might have heard about this concept. The basic idea is that the propensity score is the likelihood of being treated given a set of calories.

159
00:18:55,190 --> 00:18:59,840
So in this case, the treatment is basically receiving a phone call.

160
00:18:59,840 --> 00:19:08,510
The koalas in this case obviously would be the, let's say, individual preferences by default, how much they like music or something else.

161
00:19:08,510 --> 00:19:15,590
So the idea here is basically to pair the individuals that have a very similar propensity score such

162
00:19:15,590 --> 00:19:22,820
that we kind of mimic the treatment assignment in randomised experiments to remove selection bias.

163
00:19:22,820 --> 00:19:30,230
So typically, the way that we match people based on demographic information, age,

164
00:19:30,230 --> 00:19:37,040
gender like this stratification in this case, and we would like to explore some new options.

165
00:19:37,040 --> 00:19:47,450
Basically, what we looked at is the review of the preferences of individuals through their mobile phone choices or mobility history.

166
00:19:47,450 --> 00:19:56,510
So there is a series of review with preferences and economics that especially with business and the marketing,

167
00:19:56,510 --> 00:20:06,680
which says that your preferences is sometimes reviewed by what you do, like the products that you buy, for example.

168
00:20:06,680 --> 00:20:18,230
So in this case, we we basically look at the mobility histories, which means the frequencies of the people visiting different locations in the past.

169
00:20:18,230 --> 00:20:25,520
And we hope that the fact that they visit different locations will kind of captured their interest.

170
00:20:25,520 --> 00:20:34,460
Yeah. So is this because you didn't have information on observed demographics or this is in addition to the observed demographic?

171
00:20:34,460 --> 00:20:37,640
So, so for us, we don't have demographics here.

172
00:20:37,640 --> 00:20:41,910
That's true. Second, on demographics, they also have limitations.

173
00:20:41,910 --> 00:20:45,770
For example, they are a information. They are not dynamic.

174
00:20:45,770 --> 00:20:53,270
Uh, so uh, what we hope here is basically to see whether this kind of a more dynamic

175
00:20:53,270 --> 00:20:59,180
representation of the preferences can lead to a better matching based method.

176
00:20:59,180 --> 00:21:05,090
And I think we have partial demographic information in this case.

177
00:21:05,090 --> 00:21:11,510
Maybe gender, I'm not sure. So we did some simple comparison to see whether there was a difference or not.

178
00:21:11,510 --> 00:21:16,340
But the main motivation of using this kind of a new koala is, in some sense,

179
00:21:16,340 --> 00:21:25,530
is to actually condition the matching based on the actual behaviour of the person in the past, rather than some static information like demographics.

180
00:21:25,530 --> 00:21:30,290
And so obviously, individual mobility history is is very.

181
00:21:30,290 --> 00:21:34,210
So how do you sort of decide how much further back in the past?

182
00:21:34,210 --> 00:21:38,090
Oh yeah. So in this case, we just look at the past six months.

183
00:21:38,090 --> 00:21:47,510
So here you can see basically, we have this little cells so that they are basically like a cell towers.

184
00:21:47,510 --> 00:21:51,980
So we just look at the frequency of the president visiting different places.

185
00:21:51,980 --> 00:21:59,900
So you can see in this case, the press. And on the left, he's my explorative because he goes to different places, more or less equal frequency.

186
00:21:59,900 --> 00:22:06,690
The president on the right, he's more exploitative because he has backs back to single location that he likes.

187
00:22:06,690 --> 00:22:16,370
So this gives us an idea of how to basically link this type of behaviour,

188
00:22:16,370 --> 00:22:26,930
which is reflected in the past history to the likelihood of doing something in this case that I can attending the event.

189
00:22:26,930 --> 00:22:39,890
So but obviously, there's no good answer how much you should look back so that you capture the the kind of genuine behaviour of the person.

190
00:22:39,890 --> 00:22:49,210
In these kind of spatial things, did you focus on weekend time if the person on the left, maybe their job just takes them around?

191
00:22:49,210 --> 00:22:53,840
It's mostly weekend weekends. Yeah, it's mostly weekend mobility.

192
00:22:53,840 --> 00:22:58,720
Yeah, which we hope captures the later related activities.

193
00:22:58,720 --> 00:23:05,810
And so at the same time, we also want to study the effect of influence with respect to the communication distance.

194
00:23:05,810 --> 00:23:08,270
So I introduced the concept, the Hope Index,

195
00:23:08,270 --> 00:23:15,140
which is like how far away in some sense you are from attending because maybe you don't receive a direct information,

196
00:23:15,140 --> 00:23:21,470
but maybe you receive the information from a friend who received a phone call from the person who attended and maybe the information got propagated.

197
00:23:21,470 --> 00:23:25,460
So that's in fact the main motivation of the study.

198
00:23:25,460 --> 00:23:29,210
For that reason, we basically have one treatment group for each hope index.

199
00:23:29,210 --> 00:23:33,530
So that means for people who receive the phone calls directly from the attendee,

200
00:23:33,530 --> 00:23:38,120
we basically put them as a treatment group and then we take a we construct the

201
00:23:38,120 --> 00:23:43,010
corresponding control group from the people who are outside of the cascade,

202
00:23:43,010 --> 00:23:51,680
but conditioned on the fact that the propensity score would be similar based on their past mobility history.

203
00:23:51,680 --> 00:23:53,840
So we just pair them like that.

204
00:23:53,840 --> 00:24:00,290
And if we do that, then basically the difference in future adoption rate, we establish upper bound of social influence.

205
00:24:00,290 --> 00:24:05,810
In fact, after you control the whole monthly basis, is this.

206
00:24:05,810 --> 00:24:11,640
So wait, so this only works if you have.

207
00:24:11,640 --> 00:24:17,220
The same call records from the same company for everyone, right?

208
00:24:17,220 --> 00:24:26,540
So like if I if I call you and you call her, but her is, but she's not in our phone company, you can't see what she's doing afterwards.

209
00:24:26,540 --> 00:24:27,660
Yeah, yeah, definitely.

210
00:24:27,660 --> 00:24:35,460
So the hope is that, well, in this case, because there is only a single network provider in the country, also with high penetration rate.

211
00:24:35,460 --> 00:24:43,770
But but in a country where you have more than one network provider, then that will usually what we see is true because I think in that case,

212
00:24:43,770 --> 00:24:47,250
most of the providers would only have a market of 20 30 percent of the people.

213
00:24:47,250 --> 00:24:54,910
So it's kind of the it's larger scale than surveys, but still it's a kind of a biased in some ways.

214
00:24:54,910 --> 00:24:58,440
So this is only based in Andorra and Andorra. There's only one provider.

215
00:24:58,440 --> 00:25:02,950
So mostly mostly, yeah, we have people from France and Spain, but mostly Andorra.

216
00:25:02,950 --> 00:25:06,990
OK. Yeah. Yes. And that's it connects to that.

217
00:25:06,990 --> 00:25:16,580
To do this, people say, You know what? You can have information on my call data or, uh, so could you friend again?

218
00:25:16,580 --> 00:25:21,360
These people give permission? Oh yeah. Well, yeah.

219
00:25:21,360 --> 00:25:29,790
The assumption is that when they basically signed a mobile phone contract and then they give the permission to these kind of studies.

220
00:25:29,790 --> 00:25:38,020
But I think that's a there's a big question that we can discuss separately because that's related to to the ethical concerns and the privacy.

221
00:25:38,020 --> 00:25:39,390
Yeah. And I'm quite interested.

222
00:25:39,390 --> 00:25:47,400
I think an implicit assumption you're making is that when somebody makes a phone call, we just assume that they are talking about the concept.

223
00:25:47,400 --> 00:25:53,400
That's right, because you can never actually track the content. So you don't know if actually they talked about the event or not.

224
00:25:53,400 --> 00:25:56,310
So you're there and we are just absolutely OK.

225
00:25:56,310 --> 00:26:02,730
Yeah, which is kind of one of the limitations that we can talk in the end because you don't carry out the survey.

226
00:26:02,730 --> 00:26:07,240
Therefore, you don't know exactly what I mean. You can't track the the content of the course.

227
00:26:07,240 --> 00:26:18,060
So, yeah, uh, so it's an upper bound of social influence because obviously there are many different observe the variables that may affect adoption,

228
00:26:18,060 --> 00:26:24,500
such as like offline TV advertisement or, you know, everything that you observe on the street or something like that.

229
00:26:24,500 --> 00:26:27,570
And so, uh, so yeah.

230
00:26:27,570 --> 00:26:36,150
So here I said, I said that we we have one treatment group for each hoping that because we want to understand who has this influence, a difference.

231
00:26:36,150 --> 00:26:42,270
Uh, for for people who are of different distance from the original attendee.

232
00:26:42,270 --> 00:26:54,840
And then basically we have this uh uh and uh, in fact, a treatment effect for each hope, uh, as I just defined.

233
00:26:54,840 --> 00:26:58,890
So what we see here basically is a curve where the access is to help index.

234
00:26:58,890 --> 00:27:04,020
So that means the, uh, distance away from the initial attendee.

235
00:27:04,020 --> 00:27:11,220
The y axis is the difference is the difference in the likelihood of attending the event in the future, uh,

236
00:27:11,220 --> 00:27:21,480
divided by the likelihood of attending and by the people in the control group, which is basically estimated using a logistic regression.

237
00:27:21,480 --> 00:27:29,700
So that basically means this number here means that if you received a phone call directly from an attendee,

238
00:27:29,700 --> 00:27:38,160
then you get one hundred and fifty percent more likely to due to adopt it in the future compared to a person who who has not been influenced.

239
00:27:38,160 --> 00:27:45,060
It's interesting that we see it's obviously we see a positive effect of social influence that's kind of expected.

240
00:27:45,060 --> 00:27:50,550
Uh, but it's mostly a dramatically for the direct contact, which makes sense.

241
00:27:50,550 --> 00:27:59,190
So if we receive a direct phone call, then then then that presumably gives a strong, uh, influence.

242
00:27:59,190 --> 00:28:05,700
But we also see that this effect remains significant for people who are further away.

243
00:28:05,700 --> 00:28:10,380
So we are mostly talking about a blue curve here for the moment.

244
00:28:10,380 --> 00:28:13,560
So you see, that is it remains significant.

245
00:28:13,560 --> 00:28:27,270
Even two, three or four hops away in this case, which indicates that the influence of the initial adopter reaches far beyond the immediate circle.

246
00:28:27,270 --> 00:28:33,060
It can even influence people who they probably don't know.

247
00:28:33,060 --> 00:28:38,330
Uh, so uh, do you understand the robustness of the results?

248
00:28:38,330 --> 00:28:46,500
We also did some tests. Uh, the first is basically to do exactly the same thing, but without controlling for Homophily.

249
00:28:46,500 --> 00:28:52,260
So that means we don't match people based on the past history.

250
00:28:52,260 --> 00:28:57,450
If you do that, you observe this purple curve is a random matching.

251
00:28:57,450 --> 00:29:05,340
You see the overestimation and the overestimation is mostly due to Homophily, and this is almost twice as much.

252
00:29:05,340 --> 00:29:11,500
And this is similar to the founding in this paper, uh, in 2009.

253
00:29:11,500 --> 00:29:17,450
The other thing that we did is basically to do a shuffling test of the Hope Hop Index,

254
00:29:17,450 --> 00:29:27,430
so if the president has a direct, uh, contact, we randomly change that number from one to five or six like that.

255
00:29:27,430 --> 00:29:29,470
And after that, we again do the magic.

256
00:29:29,470 --> 00:29:38,350
So the purpose of doing this is basically to see whether some of the variables were related to the decay patterns that we see.

257
00:29:38,350 --> 00:29:41,950
But the red curve basically shows that there is no such pattern.

258
00:29:41,950 --> 00:29:49,810
So the observed results is not likely to be mainly driven by something that we don't observe.

259
00:29:49,810 --> 00:29:59,920
So to understand how these patterns decay, so what is the mechanism that lead to this decay effect of social inference?

260
00:29:59,920 --> 00:30:03,880
We do the simple Bayesian model to to to understand it.

261
00:30:03,880 --> 00:30:11,510
So the basic idea is that in a communication cascade, an individual will basically updates for estimation of the product.

262
00:30:11,510 --> 00:30:15,910
So the product is defined that broadly, so whatever that you want to adopt,

263
00:30:15,910 --> 00:30:21,880
it can be a new app or it can be a piece of music or whether you want to attend the event or not.

264
00:30:21,880 --> 00:30:26,560
So an individual will update our estimation of the product characteristics.

265
00:30:26,560 --> 00:30:38,380
So what is the event like? Things like that after they talk to you, the people around him or her.

266
00:30:38,380 --> 00:30:44,950
And then after that, they will dynamically update the estimates by aggregating information from the neighbours.

267
00:30:44,950 --> 00:30:52,930
So then we assume that this estimation, together with the president's own preference by default,

268
00:30:52,930 --> 00:30:56,890
will basically form an evaluation of of the person for the event.

269
00:30:56,890 --> 00:30:59,080
So if the present estimates,

270
00:30:59,080 --> 00:31:10,780
the event is kind of like a pop music event with a huge crowd and his own preference is basically he likes these large scale events.

271
00:31:10,780 --> 00:31:19,420
And then he put together these two pieces of information to form his own evaluation, which is this is the event that I like to go.

272
00:31:19,420 --> 00:31:25,550
And then he will basically make the adoption decision following a simple binary probability distribution.

273
00:31:25,550 --> 00:31:30,430
And so the estimation is basically done like here.

274
00:31:30,430 --> 00:31:33,280
So Greg basically received a phone call from Bob,

275
00:31:33,280 --> 00:31:42,160
and then he will basically updated his estimation based on the evaluation of Bob and the fact that he knows that Bob likes certain things.

276
00:31:42,160 --> 00:31:45,940
And then basically, this will keep going, keep going like that.

277
00:31:45,940 --> 00:31:55,000
And and then after a while, we could basically see the estimation of the event by all the individuals.

278
00:31:55,000 --> 00:32:04,420
And now we can see how this estimation basically change when we move further away from the initial adopter.

279
00:32:04,420 --> 00:32:12,940
So what we measure here is the difference in the estimation between the individuals who are in the Cascades and the initial adopter.

280
00:32:12,940 --> 00:32:20,140
So the x axis here is again the distance, the hope index. The Y is basically the mean squared difference between the evaluation.

281
00:32:20,140 --> 00:32:25,720
What you see here is that this difference increases as the Hope Index increases.

282
00:32:25,720 --> 00:32:38,890
So this could mean the estimation of the event become less accurate and contains mobiles because the estimation of individuals contain bias.

283
00:32:38,890 --> 00:32:42,580
And this bias can accumulate along the information cascade.

284
00:32:42,580 --> 00:32:52,090
And also the fact that if you receive inference from multiple people can also have this aggregation effect.

285
00:32:52,090 --> 00:32:58,270
So in the end, people here they have a less accurate estimation of the event,

286
00:32:58,270 --> 00:33:01,930
which might be related to the decline pattern of the inference in this case.

287
00:33:01,930 --> 00:33:12,460
But this is still something that we're wanting to investigate further. So it's kind of like a conjecture that we hope to test further.

288
00:33:12,460 --> 00:33:26,590
So the main findings in this study is that first, we we try to use the revealed preferences instead of static demographics as covariance,

289
00:33:26,590 --> 00:33:32,950
uh, to, uh, quantify the inference in matching based estimation.

290
00:33:32,950 --> 00:33:40,780
Uh, and and we think this can actually overcome a lot of difficulties in traditional our city based approaches because,

291
00:33:40,780 --> 00:33:48,850
uh, sometimes, uh, for privacy reasons, you don't have the corroborates as you like.

292
00:33:48,850 --> 00:33:54,220
Sometimes you have, but they remain static, so you can't capture the dynamic changes.

293
00:33:54,220 --> 00:34:01,720
If there are, uh, the empirical finding is that we observed these long range effects of social influence, uh,

294
00:34:01,720 --> 00:34:11,090
which may have implications in scenarios like our marketing or public health management when you try to basically use a.

295
00:34:11,090 --> 00:34:16,490
Online communication to influence people's decision making,

296
00:34:16,490 --> 00:34:24,680
all these situations and the limitations are mostly the identification of the initial

297
00:34:24,680 --> 00:34:29,250
adopters is not precise because we don't really know whether the person at hand or not.

298
00:34:29,250 --> 00:34:43,100
We just know that the person had some activities within the very small range from the venue of the event, but we didn't actually ask the people.

299
00:34:43,100 --> 00:34:50,240
So we don't know whether they actually attended or they just passed by Keith.

300
00:34:50,240 --> 00:34:58,770
So to come back to this point about matching was on on static demographics versus the more dynamic ones dynamic characteristics,

301
00:34:58,770 --> 00:35:05,730
I mean, to some extent, dynamic attributes, there's a certain androgyny do there because things that are changing are they

302
00:35:05,730 --> 00:35:10,050
could be changing because people are themselves being influenced themselves.

303
00:35:10,050 --> 00:35:14,220
And we don't really observe that experience. So how do we?

304
00:35:14,220 --> 00:35:16,200
So while I'm sympathetic to your point,

305
00:35:16,200 --> 00:35:22,740
I also kind of wonder about that issue of indulgent idea of the fact that mobility itself reflects influence processes that are happening.

306
00:35:22,740 --> 00:35:28,590
I go to this museum where someone told me to go to. Yeah, I think that's true.

307
00:35:28,590 --> 00:35:36,360
I didn't present a result where we basically compare the two different matching wines based on demographics.

308
00:35:36,360 --> 00:35:40,620
The other is based on demographics plus the mobility history.

309
00:35:40,620 --> 00:35:49,650
And the results are mostly similar, which makes us think like the mobility histories contain the information of the demographics in some sense.

310
00:35:49,650 --> 00:35:52,440
But this by itself is a very interesting question to ask,

311
00:35:52,440 --> 00:36:01,390
like whether these observe the behaviour to which extent they captured the static attributes that we use in the traditional case.

312
00:36:01,390 --> 00:36:09,900
But more importantly, if you are not allowed to collect these kind of demographics, then then then these can be more useful.

313
00:36:09,900 --> 00:36:15,820
And other obvious things is that the dramatic exists multiple treatment effect where you receive multiple phone calls.

314
00:36:15,820 --> 00:36:22,240
But right now, the distance that we define is basically the shortest distance to the attendees.

315
00:36:22,240 --> 00:36:27,000
So. So obviously, that's something that we can do further.

316
00:36:27,000 --> 00:36:34,500
But the main idea is that given the social network in this case, we want to see how this influence payroll and the how it can propagate, basically.

317
00:36:34,500 --> 00:36:39,180
But in some cases, you don't observe this communication network.

318
00:36:39,180 --> 00:36:47,020
But we just observe that the adoption behaviour of people so they decide to do something if.

319
00:36:47,020 --> 00:36:56,480
Yes. So could you speak a bit more about how the estimate of the product was done because you calculated it means great difference?

320
00:36:56,480 --> 00:37:01,250
So I guess these estimates were actual numbers and what went into the calculation of these?

321
00:37:01,250 --> 00:37:06,200
Yeah, so so do we initialised as following a certain distribution.

322
00:37:06,200 --> 00:37:16,640
And then after that, let's say, if you called me, then basically you will have your estimation, which leads to your your evaluation.

323
00:37:16,640 --> 00:37:21,770
And we also assume that each person has the initial vector of preferences.

324
00:37:21,770 --> 00:37:30,020
So based on what I receive from you, I will basically just use a baseline rule to compute the posterior of my estimation,

325
00:37:30,020 --> 00:37:34,190
conditioned on the fact that I observe the information from you also.

326
00:37:34,190 --> 00:37:39,080
So it's basically the posterior is basically the prior proportional to the prior times.

327
00:37:39,080 --> 00:37:48,020
The likelihood the prior is basically my original sorry, the prior is basically the estimation from the past timestamp.

328
00:37:48,020 --> 00:37:59,840
The likelihood is basically what is the likelihood of this given estimation, such that it leads to the information that I received from you?

329
00:37:59,840 --> 00:38:11,030
So you just a simulate all these quantities by putting distributions on top of them and then you just compute.

330
00:38:11,030 --> 00:38:16,520
This estimate is sort of an average of the vector is it's a vector.

331
00:38:16,520 --> 00:38:25,160
Is it a vector? Yeah. Yeah, OK. And then this is basically a square that is just a difference between the vector.

332
00:38:25,160 --> 00:38:33,980
Yeah. So the second question that we want to ask, which is just kind of conceptually the reverse of the problem,

333
00:38:33,980 --> 00:38:43,280
if you observe some and adoption actions or whatever actions, what could be the opposite of the social network that lets you these actions?

334
00:38:43,280 --> 00:38:48,620
We will use a slightly different framework here, which is based on game theory.

335
00:38:48,620 --> 00:38:58,370
So we are going to make a small transition here. But the conceptual, uh, clustering is the reverse of the first one that we ask.

336
00:38:58,370 --> 00:39:04,550
So let's start with some simple examples consider a group of students making choices on an educational effort.

337
00:39:04,550 --> 00:39:12,500
So how much time you use? You want to put just two to a certain study, so they will basically follow a number of rules.

338
00:39:12,500 --> 00:39:19,100
The first rule will be the the making effort is costly, obviously for me, but I will benefit from my own efforts, obviously.

339
00:39:19,100 --> 00:39:23,960
And in this case, I also benefit from my friend's effort because it's a collective study.

340
00:39:23,960 --> 00:39:31,970
It's, of course, projects like that. So if we have these kind of rules, then the president will tend to put more effort.

341
00:39:31,970 --> 00:39:42,920
The friends of the president put more effort. So one scenario could be this where the colours divide how much effort they want to put in.

342
00:39:42,920 --> 00:39:53,020
So you see the president here, it was a lot of effort because many of his friends put a lot of effort, and the opposite is true for this blue guy and.

343
00:39:53,020 --> 00:40:01,810
A second example is to consider a group of students making choice to buy a book, whether each individual wants to buy a book or not.

344
00:40:01,810 --> 00:40:04,870
Obviously, in this case, buying a book is costly.

345
00:40:04,870 --> 00:40:13,510
I cost the time and money, so if a friend of mine will buy, then I will not buy based on the assumption that I can easily borrow.

346
00:40:13,510 --> 00:40:18,250
But even now, my friends will abide by it because I definitely need the book.

347
00:40:18,250 --> 00:40:25,660
So if we have this kind of rules, then we are likely to end up in a situation like these.

348
00:40:25,660 --> 00:40:30,640
The right guys decide to buy a book because none of their neighbours decides to

349
00:40:30,640 --> 00:40:36,580
buy and the Brueghel they don't buy because at least one neighbour about the book.

350
00:40:36,580 --> 00:40:39,880
So this is another scenario.

351
00:40:39,880 --> 00:40:48,970
Both of the scenarios you can see that the existence of strategic interactions between people, which can be modelled as games.

352
00:40:48,970 --> 00:40:55,030
Game theory is basically a mathematical models of a rational decision making between individuals,

353
00:40:55,030 --> 00:41:03,850
and usually it contains key concepts such as a player actions and actions kind of binary, whether by a book or not.

354
00:41:03,850 --> 00:41:13,390
It can also be continuous, like the amount of effort that you put, and the payoff is basically how much reward you got for your action.

355
00:41:13,390 --> 00:41:21,640
What is so all the these concepts they remain in what we call on game social networks,

356
00:41:21,640 --> 00:41:29,710
but for the concepts here we have an additional information, which is the interaction network.

357
00:41:29,710 --> 00:41:37,870
So friendship network, for example. So the interesting idea here would be to see how they're in.

358
00:41:37,870 --> 00:41:42,940
The virtual actions can be related to the structure of the network.

359
00:41:42,940 --> 00:41:48,700
So I can give you an example. In the previous cases, if you are central in the network, you are well connected,

360
00:41:48,700 --> 00:41:57,310
then obviously that will affect how much effort that you're put in projects or whether you are going to buy a book or not.

361
00:41:57,310 --> 00:42:02,020
So the relationship between the individual decision making and the structure of the

362
00:42:02,020 --> 00:42:07,890
network is the most interesting aspect in these models of the games on networks.

363
00:42:07,890 --> 00:42:17,560
And obviously, the basic assumption is that the payoff of one individual depends on her action, but also the actions of the neighbours.

364
00:42:17,560 --> 00:42:26,650
So the two things that we talked about are actually examples of what we call strategic accomplishments and substitute.

365
00:42:26,650 --> 00:42:39,280
So I think this is quite a self-explanatory. So there are many studies dedicated to this feud, which is a subject of economics.

366
00:42:39,280 --> 00:42:47,980
And if not, people look at the equilibrium of games played on networks and on a given or preferred network.

367
00:42:47,980 --> 00:42:52,900
They're mostly interested in how action and the payoff depends on the network structure.

368
00:42:52,900 --> 00:43:00,020
Uh, in computer science, the same ideas is studied under the name of graphical games.

369
00:43:00,020 --> 00:43:08,570
Uh, what they're interested in is mostly to develop algorithms to compute the equilibrium in the computational sense.

370
00:43:08,570 --> 00:43:18,190
And so what we have more interested in this work is a reverse problem, which is trying to infer the network given the conditions that we observe.

371
00:43:18,190 --> 00:43:20,530
So we don't assume that we have a network,

372
00:43:20,530 --> 00:43:26,800
but we assume that the people play games in the hidden network and then we want to basically recover that structure.

373
00:43:26,800 --> 00:43:32,290
You can find a lot of examples. We observe individual decisions, adoptions, but not social networks.

374
00:43:32,290 --> 00:43:35,710
We observed the research and development activities of firms,

375
00:43:35,710 --> 00:43:45,430
but we don't observe the collaboration network of firms or policies of countries, but not the political alliances.

376
00:43:45,430 --> 00:43:54,160
So obviously, the the difficulty here is the complexity of the network, which makes the inference more difficult.

377
00:43:54,160 --> 00:43:59,680
As a result, we have to kind of put some structure to the game.

378
00:43:59,680 --> 00:44:06,040
Otherwise, it would be very difficult to make any inference the specific structure.

379
00:44:06,040 --> 00:44:14,800
I mean, the the broad definition could be the payoff of a player UI is basically a function of his own action A.I.

380
00:44:14,800 --> 00:44:20,470
Which are the actions of his neighbours in the observer network or in this case,

381
00:44:20,470 --> 00:44:25,000
we can say is observe the network and then the connectivity, basically.

382
00:44:25,000 --> 00:44:31,360
So one particular formulation of the payoff is this linear quadratic model.

383
00:44:31,360 --> 00:44:33,070
It's quite easy to understand.

384
00:44:33,070 --> 00:44:42,880
Here we have individual actions A.I. and then we have this marginal benefit by which basically decides independent from independent from the others.

385
00:44:42,880 --> 00:44:50,800
If I increase my action, how much of a benefit I would get. And then we have this network factor beta in front of this term.

386
00:44:50,800 --> 00:44:59,070
So this time we can basically. Think about is the eight times the summation of the neighbours action,

387
00:44:59,070 --> 00:45:03,270
so you can clearly see that the first two terms corresponds to a kind of individual effort.

388
00:45:03,270 --> 00:45:10,450
This is the individual benefits this is the cost is an active squad's time and then this training is basically the network effect.

389
00:45:10,450 --> 00:45:18,240
So if you have a lot of people who are your neighbours and they they all have a very high age,

390
00:45:18,240 --> 00:45:24,090
then you know, if you put a higher rate than it would increase your own utility.

391
00:45:24,090 --> 00:45:29,400
So that's basically one specific model that we use.

392
00:45:29,400 --> 00:45:35,400
And why do we choose this specific function, payroll function first?

393
00:45:35,400 --> 00:45:38,010
It's a continuous. It can capture continuous actions.

394
00:45:38,010 --> 00:45:46,440
That aid doesn't have to be discrete or in many cases, the A's are assumed to be discrete in the literature.

395
00:45:46,440 --> 00:45:52,170
Also, it's very easy for us to capture both strategic accomplishments and the substitute.

396
00:45:52,170 --> 00:45:56,940
You just need to choose a positive beta or negative beta, the negative beta,

397
00:45:56,940 --> 00:46:06,180
which basically mean that there is a kind of a negative network effect and this is a simple quadratic model,

398
00:46:06,180 --> 00:46:11,220
but it can also be used to approximate the more complex nonlinear payoffs.

399
00:46:11,220 --> 00:46:23,760
So that's also makes it general. And this is also that while adopting the model in the literature of games on networks again,

400
00:46:23,760 --> 00:46:29,370
we have a lot of examples that I just gave before that these games can think about educational

401
00:46:29,370 --> 00:46:37,860
efforts or collaboration or even individual mobility choices in the context of urban dynamics.

402
00:46:37,860 --> 00:46:48,000
So let's focus on the most important advantage of this specific model in game theory were mostly interesting equilibrium.

403
00:46:48,000 --> 00:47:00,150
So what is the kind of a stable situation where a person will not tend to make a change of his action, if not all of his neighbours, which ends?

404
00:47:00,150 --> 00:47:07,200
So the concept of national equilibrium is to capture this stable situation.

405
00:47:07,200 --> 00:47:10,770
In this case, the equilibrium is easy to derive.

406
00:47:10,770 --> 00:47:15,300
We can simply take it the first order condition of stability,

407
00:47:15,300 --> 00:47:24,240
so we take the derivative of this utility with respect to the individual action, and then we end up in the equation here.

408
00:47:24,240 --> 00:47:30,780
So B here is basically a vector of the marginal benefit if you remember the A's of active actions.

409
00:47:30,780 --> 00:47:36,090
So you see that the actions A's basically the inverse of this matrix times B.

410
00:47:36,090 --> 00:47:45,360
And so we need a few assumptions to to guarantee that this inversion exists,

411
00:47:45,360 --> 00:47:49,800
which is basically the spectral radius of this matrix is smaller than one.

412
00:47:49,800 --> 00:47:58,140
And the spectral radius is basically the largest single the value of the of the matrix.

413
00:47:58,140 --> 00:48:00,900
If that's the case, then you can guaranteed a matrix.

414
00:48:00,900 --> 00:48:08,730
Inversion exists, and the equilibrium is unique and unstable, so that's just a technical conditions.

415
00:48:08,730 --> 00:48:14,970
But if we observe this equilibrium more carefully, we can we can find something interesting.

416
00:48:14,970 --> 00:48:25,520
This inversion can be rewritten as even if some of the terms here, so be it.

417
00:48:25,520 --> 00:48:28,530
And this is closely related to a concept in network science,

418
00:48:28,530 --> 00:48:40,290
which is called the voltage centrality that just counts the total number of works from one single person to anybody in the network.

419
00:48:40,290 --> 00:48:47,640
So and the interesting thing here is that if you look at the payload function here,

420
00:48:47,640 --> 00:48:52,980
it looks like this pair of interdependency only existing, well, hot neighbour.

421
00:48:52,980 --> 00:48:57,180
But if you look at the equation here, actually, you see that under the equilibrium,

422
00:48:57,180 --> 00:49:03,990
this pair of dependency spreads indirectly through the whole network, which is something interesting to observe.

423
00:49:03,990 --> 00:49:09,720
And you can also have some other interpretation far more signal processing viewpoint.

424
00:49:09,720 --> 00:49:16,770
But I will skip that because probably that's not very easy to explain in the short period of time.

425
00:49:16,770 --> 00:49:24,570
But under this model, we basically come up with a learning algorithm that helps us to basically infer the G from the air.

426
00:49:24,570 --> 00:49:33,510
So that's the the the idea that we had from the beginning, and we we basically consider K independent games.

427
00:49:33,510 --> 00:49:41,160
All of them have different marginal benefit vectors, so you can think about the adoption of different types of products.

428
00:49:41,160 --> 00:49:44,190
The marginal benefits for each of them will be different for individuals.

429
00:49:44,190 --> 00:49:49,680
Therefore, the equilibrium actions will be different, but the network is assumed to be the same.

430
00:49:49,680 --> 00:50:05,060
So we. Try to infer the network structure G and together with the marginal benefit to be given only the Action A and the parameter beta.

431
00:50:05,060 --> 00:50:09,260
What we do is to solve this following optimisation problem.

432
00:50:09,260 --> 00:50:17,300
The first to here we can see is just a condition that is matched when the equilibrium of the action is observed.

433
00:50:17,300 --> 00:50:25,430
So you want to minimise this term so that it basically means that your actions close to equilibrium, then we want to put some other conditions on.

434
00:50:25,430 --> 00:50:30,830
The variable that we want to optimise for the G is basically the network structure.

435
00:50:30,830 --> 00:50:40,460
So we want it to be symmetric non-active and then also constraint on the total sum of the weights of the edges.

436
00:50:40,460 --> 00:50:46,550
This just guarantees that the variable that we solve in the end is a valid graph topology, a structure.

437
00:50:46,550 --> 00:50:54,620
We also need some constraint or prior information on this, be the marginal benefits.

438
00:50:54,620 --> 00:51:02,780
The simple thing that we consider is, again, an idea taken from this idea of Homophily.

439
00:51:02,780 --> 00:51:10,190
So two people are likely to form a link if they have similar characteristics.

440
00:51:10,190 --> 00:51:15,080
This characteristic here is, in fact, that's the marginal benefit. So you like music?

441
00:51:15,080 --> 00:51:19,970
Therefore, taking certain actions will lead to a higher power, for example.

442
00:51:19,970 --> 00:51:27,350
So if we think of this, uh, marginal benefit as a single value, let's say, instead of a vector,

443
00:51:27,350 --> 00:51:33,110
we just think of either like a single value and then we represent that value as a bar in a network.

444
00:51:33,110 --> 00:51:40,010
So the the the bars that are upwards are positive, the blue ones, they are negative.

445
00:51:40,010 --> 00:51:48,590
So in this case, because two people will form a likely to form a link if they are marginal, benefits are similar.

446
00:51:48,590 --> 00:51:54,590
Therefore, if you look at this information, which is the marginal benefits on top of this network,

447
00:51:54,590 --> 00:51:59,630
you will see that the transitions of these values will be kind of smooth across the edges because

448
00:51:59,630 --> 00:52:05,420
that's based on the assumption that a link will only form between people who have similar attributes.

449
00:52:05,420 --> 00:52:14,780
So in computer science, you can use a term which is this thing based on what is called a passive matrix of the network to quantify.

450
00:52:14,780 --> 00:52:24,260
This is basically say that if there is a connexion which is non-zero, then I want beyond Vee-Jay to be similar.

451
00:52:24,260 --> 00:52:35,120
So in our context, we use this term here, which basically means we assume a kind of a homogenous distribution of the marginal benefits.

452
00:52:35,120 --> 00:52:41,300
And then after we have this term, we can basically solve this problem, uh,

453
00:52:41,300 --> 00:52:48,030
using specific computational methods, uh, which I will probably not talk too much.

454
00:52:48,030 --> 00:52:57,740
And basically, it's you have this adjointe, a variable GMB, which leads to a non convex optimisation,

455
00:52:57,740 --> 00:53:02,730
but you can use the alternating minimisation to solve all both of them iteratively.

456
00:53:02,730 --> 00:53:10,550
If I'm just, like, probably over oversimplify a lot of this, but just because not yet, anyways.

457
00:53:10,550 --> 00:53:18,860
So part of it is that you have the same network, but your training algorithm, which is there's a variation in the marginal benefit of different.

458
00:53:18,860 --> 00:53:25,400
Yeah, different. Yeah, that's right. So you're saying I have the same network many, many times, but yeah.

459
00:53:25,400 --> 00:53:29,440
Yeah. And then I'm trying to guess a person's position. Network.

460
00:53:29,440 --> 00:53:34,580
Yeah, that's right. Yeah. It can be extended to cases where networks are changing.

461
00:53:34,580 --> 00:53:38,960
But right now, we just consider six to topology, but a different marginal benefits.

462
00:53:38,960 --> 00:53:44,300
And did you put any constraints on the different types of networks, like the topology of the networks?

463
00:53:44,300 --> 00:53:48,500
That's a very good question. Right now, we don't because this constraint is very general.

464
00:53:48,500 --> 00:53:52,530
It's the basic constraint that guarantees a value, the network topology.

465
00:53:52,530 --> 00:53:56,390
You can put things like a degree distribution or other things,

466
00:53:56,390 --> 00:54:02,270
which we studied in the experiments, but now we don't put them in the constraint, the explicitly.

467
00:54:02,270 --> 00:54:09,280
You can if you have some, you know, prior information about something and then you try to put that as a constraint.

468
00:54:09,280 --> 00:54:19,670
Uh, the basic intuition is that if this spectrum radius role Beta G is zero, then obviously, you know,

469
00:54:19,670 --> 00:54:24,740
you don't have any information of the network from a because you just got a minus B, right?

470
00:54:24,740 --> 00:54:30,320
So so there is no hope that you can infer the G from from that.

471
00:54:30,320 --> 00:54:37,880
But if this a b g becomes a bigger and bigger, then obviously a will containing the information from the network.

472
00:54:37,880 --> 00:54:42,350
The extreme case, if Peter G has a radius of one,

473
00:54:42,350 --> 00:54:48,590
then we can work out that this action is proportional to the eigenvectors ontology of the network and that eigenvectors.

474
00:54:48,590 --> 00:54:52,060
Ontology, in fact, has a strong information about the.

475
00:54:52,060 --> 00:55:04,930
Activity of the network, so that that's kind of an intuition, why from action a we can go back to to get to the topology, see.

476
00:55:04,930 --> 00:55:10,480
So we tested this idea in both the synthetic and real world, the data.

477
00:55:10,480 --> 00:55:14,320
So these are basically the different types of pathologies that you mentioned.

478
00:55:14,320 --> 00:55:19,840
These are random graph models that are typically studied in network science.

479
00:55:19,840 --> 00:55:30,250
They basically capture different types of activities. And what we do is basically to first to generate a grant to run the network.

480
00:55:30,250 --> 00:55:41,020
Then we basically use this idea to compute an action based on a simulation of the amorphous marginal benefits,

481
00:55:41,020 --> 00:55:44,860
which is assumed to be a Gaussian distribution here. So we got the G.

482
00:55:44,860 --> 00:55:49,540
We compute the B following these and then we got a.

483
00:55:49,540 --> 00:55:57,610
And now we just use the algorithm to infer GMB from a that's the basic procedure.

484
00:55:57,610 --> 00:56:06,430
And then we basically compare it with some other, like a well known methods in computer science, mostly, and they evaluate the performance.

485
00:56:06,430 --> 00:56:12,760
And this is one set of experiments that we do. Each figure is basically one type of random graph.

486
00:56:12,760 --> 00:56:20,530
What we observe here are the x axis is the spectrum radius of the Beta G matrix, and the Y is the evaluation,

487
00:56:20,530 --> 00:56:24,760
which is the similarity between what we infer and the ground shows that we start with.

488
00:56:24,760 --> 00:56:33,630
So you can see that when we increase the spectral radius, the performance increases, which follows the intuition that I just explained.

489
00:56:33,630 --> 00:56:47,100
And then. The method that we have here in general achieves a better performance than a simple baseline, which is sample correlation.

490
00:56:47,100 --> 00:56:52,230
And also another well-known method in computer science, which is called the graphical lasso.

491
00:56:52,230 --> 00:56:53,910
I'm not sure if you're familiar with sound,

492
00:56:53,910 --> 00:57:03,660
but the idea is to see whether this new formulation can actually improve the performance compared to some existing models.

493
00:57:03,660 --> 00:57:09,750
The more interesting question in this case would be whether the algorithm works equally well three and four different types of graphs,

494
00:57:09,750 --> 00:57:17,730
like graphs of different kind of activity. So to do that, basically, we test it for these three different types of graph.

495
00:57:17,730 --> 00:57:22,500
Each of these running graph models have parameters to decide that the densities of the edges.

496
00:57:22,500 --> 00:57:32,310
So, for example, in others or any network, edges are created independently with a given probability distribution.

497
00:57:32,310 --> 00:57:38,160
So if the P is high, then the network becomes dense.

498
00:57:38,160 --> 00:57:43,650
Another example would be this Barbash Aabot network,

499
00:57:43,650 --> 00:57:53,190
which is a generative model that each time you just add a new vertex and that connects the new vertex to the existing nodes in the graph.

500
00:57:53,190 --> 00:57:57,360
So if you decide every time you put new nodes,

501
00:57:57,360 --> 00:58:02,130
that means every time you just add new edges and the rule is that you only

502
00:58:02,130 --> 00:58:06,900
connect yourself at the new nodes to the notes that are already well connected.

503
00:58:06,900 --> 00:58:12,180
So that in the end, you have these preferential attachment.

504
00:58:12,180 --> 00:58:13,500
So they are just different networks.

505
00:58:13,500 --> 00:58:21,630
But what we observed here is that when these parameters are high, which leads to denser networks, the performance drops.

506
00:58:21,630 --> 00:58:25,270
And one reason for that could be if you have a very dense network,

507
00:58:25,270 --> 00:58:32,400
then the Connexions basically have this influence, which are kind of mingled against with each other.

508
00:58:32,400 --> 00:58:38,820
So it becomes very difficult for you to recover from just the observed actions.

509
00:58:38,820 --> 00:58:46,890
If the network is sparse, then the likely correlation between the actions indicates more the existence of a link.

510
00:58:46,890 --> 00:58:55,620
So that's the basic intuition. We also did some real world data to infer so.

511
00:58:55,620 --> 00:58:56,130
In this case,

512
00:58:56,130 --> 00:59:03,130
it's an inference of a social network structure where the data basically are the almost two hundred households in a village in rural India.

513
00:59:03,130 --> 00:59:13,230
This is one classical study economics and the games in this case are basically the number of facilities adopted by each household.

514
00:59:13,230 --> 00:59:21,630
And so it's arguably a and strategic accompaniments game because of the conformity to social norms,

515
00:59:21,630 --> 00:59:29,550
which basically lead to higher benefits in the village. And then we basically take the data, which is the actions,

516
00:59:29,550 --> 00:59:37,020
and then gets the social network and to compare that with self-reported the social network to see how it works.

517
00:59:37,020 --> 00:59:42,570
So in this case, we can get a good performance better than the baseline.

518
00:59:42,570 --> 00:59:49,350
The other example is a dream for a trade network. So that's a similar idea.

519
00:59:49,350 --> 00:59:56,520
But the slight the difference is that in this case, it's more like a strategic substitute instead of a compliment.

520
00:59:56,520 --> 01:00:02,970
But the basic idea is to observe actions which are important Oxford X worth of countries,

521
01:00:02,970 --> 01:00:09,540
and we want to see whether they're looking for help us to infer a trade network that we don't observe.

522
01:00:09,540 --> 01:00:19,530
So that's basically a high level idea. So again, here, I think the interesting thing would be the application in the practical scenarios.

523
01:00:19,530 --> 01:00:24,840
If you only observe the actions by nodes, the network, then obviously if you can infer the network,

524
01:00:24,840 --> 01:00:29,610
you can use it to detect the communities, which is useful for stratification.

525
01:00:29,610 --> 01:00:36,420
You can compute the centrality measures which can help you to design targeting interventions,

526
01:00:36,420 --> 01:00:42,760
and you can basically achieve planning objectives by looking at the network.

527
01:00:42,760 --> 01:00:48,240
So the two examples here could be if you want to maximise the total utility of the players,

528
01:00:48,240 --> 01:00:54,810
then you can adjust the marginal benefits that you learnt, uh,

529
01:00:54,810 --> 01:00:57,330
so that it is proportional to the eigenvectors ontology,

530
01:00:57,330 --> 01:01:04,470
which is proving this paper you can think about in reducing inequality between different players like different communities in the city.

531
01:01:04,470 --> 01:01:16,020
Uh, by adjusting the connexions between the different communities after you interact so you can encourage new links and discourage existing links.

532
01:01:16,020 --> 01:01:18,270
There are several open issues here.

533
01:01:18,270 --> 01:01:26,460
The determination of the parameter data, um, the theoretical understanding of the algorithm when it works with what guarantee,

534
01:01:26,460 --> 01:01:35,680
that's more like a, uh, a guarantee that we would like to have so that we can actually apply it in a real scenario as.

535
01:01:35,680 --> 01:01:42,190
Then obviously other payroll functions or applications.

536
01:01:42,190 --> 01:01:50,230
So the high level idea here is kind of opposite to the first problem that I talked if we observe certain actions.

537
01:01:50,230 --> 01:01:56,140
Can we try to basically go back to infer the hidden links between them?

538
01:01:56,140 --> 01:02:02,860
So these two basically combine together and in a sense,

539
01:02:02,860 --> 01:02:08,650
that I think collective decision making is key to understand the behaviour and the

540
01:02:08,650 --> 01:02:14,150
network based on learning can provide some natural models and to us to do that.

541
01:02:14,150 --> 01:02:24,460
The caveat here is more like the learning and inference they need to be done in combination with social series for better interpretability.

542
01:02:24,460 --> 01:02:32,710
Otherwise, they just become a computer science exercise without the implications in policymaking.

543
01:02:32,710 --> 01:02:40,390
More importantly, you need to think about the ethical implications and consequences when you build models to understand human behaviour.

544
01:02:40,390 --> 01:02:47,200
It might lead to marginalisation of some disadvantaged groups in some sense or situation, so.

545
01:02:47,200 --> 01:02:53,710
And you need to be basically extra careful when you make these modelling assumptions,

546
01:02:53,710 --> 01:03:00,140
but that basically needs help from social scientists to work in a more collaborative setting.

547
01:03:00,140 --> 01:03:06,820
And so finally, there are a few papers and the co-authors of this paper, so we start out and thank you for your attention.

548
01:03:06,820 --> 01:03:21,610
Take the question. Oh, you can trade off that.

549
01:03:21,610 --> 01:03:33,220
And one of my well, just for clarity sake, if you wanted to infer the the the location of someone in the network, do you using your method,

550
01:03:33,220 --> 01:03:40,450
you need to know what the network like topology is first to be able to identify where they're in that point,

551
01:03:40,450 --> 01:03:44,680
you just because your gains held topology constant. Right?

552
01:03:44,680 --> 01:03:48,490
No, I think you just need to to observe these actions, which are vectors.

553
01:03:48,490 --> 01:03:52,720
OK. And then basically, you get what is called adjacency matrix, right?

554
01:03:52,720 --> 01:03:59,950
And that addresses a matrix supposedly captures the kind of activity pattern. So you don't need the you don't need the network topology eventually.

555
01:03:59,950 --> 01:04:06,160
Yeah, that's basically what you want to infer. So who is just one individual location, but the entire topology of the network?

556
01:04:06,160 --> 01:04:08,470
So yeah, so that in some sense,

557
01:04:08,470 --> 01:04:15,190
I think the location of the network is doesn't matter that much because you can redraw the network in whatever way, right?

558
01:04:15,190 --> 01:04:22,510
So so they just I think they depend on the specific layout of the graph, if you like.

559
01:04:22,510 --> 01:04:27,640
But but the information is basically by the A's connected to be so whether there

560
01:04:27,640 --> 01:04:31,990
is the existence of that link that's completely captured by the adjacency matrix.

561
01:04:31,990 --> 01:04:36,620
So we infer that just from the electoral observation of actions.

562
01:04:36,620 --> 01:04:37,150
OK.

563
01:04:37,150 --> 01:04:46,330
And just following up on that, a lot of what we've been discussing lately is that with digital trace data, we we don't know what we're missing so far.

564
01:04:46,330 --> 01:04:54,910
And this method seems like it could be one way to start getting at that in the context of social network data.

565
01:04:54,910 --> 01:05:01,270
We could try. Yeah. Although this particular matter is based on the assumption of network games,

566
01:05:01,270 --> 01:05:06,910
which means that the actions people take are kind of based on strategic interactions.

567
01:05:06,910 --> 01:05:14,620
Some abilities, whether they. So there is an interesting example in urban dynamics that can feed this,

568
01:05:14,620 --> 01:05:20,710
which is a think about artificial city, where we only have two locations wines to centre.

569
01:05:20,710 --> 01:05:22,780
The other is the periphery.

570
01:05:22,780 --> 01:05:30,820
Now everybody needs to basically decide how many times they go to the city centre for social benefits, like by interacting with down there.

571
01:05:30,820 --> 01:05:40,480
So in that case, the number of visits is basically the action. And then you obviously want to visit more if you are central in the social

572
01:05:40,480 --> 01:05:44,170
network because there is a high likelihood that you can target with more people.

573
01:05:44,170 --> 01:05:52,990
So in that specific scenario, you can think of inferences the hidden social natural connexion from the action,

574
01:05:52,990 --> 01:05:58,060
which is the number of visits to page to the city centre.

575
01:05:58,060 --> 01:06:05,020
I guess if the digital traces can fit into something like that, like the fact that I move to a certain place,

576
01:06:05,020 --> 01:06:10,000
this is likely to be affected by the fact that you are so gorgeous in place.

577
01:06:10,000 --> 01:06:18,980
Then then it might be possible to do this because my action we're not just dependent on my own action, but also you actually let it make sense.

578
01:06:18,980 --> 01:06:26,400
Yeah. All right.

579
01:06:26,400 --> 01:06:35,430
Thank you for the talk. I was wondering about the first beach we're talking about on the network through telephone

580
01:06:35,430 --> 01:06:43,500
if where you could insert some alternative demographics by just looking at how people call.

581
01:06:43,500 --> 01:06:46,720
Yeah, yeah. Well, I think there are people who do that.

582
01:06:46,720 --> 01:06:55,260
Uh, I have some colleagues who try to use communication patterns, mostly in what is called the ego network.

583
01:06:55,260 --> 01:07:02,240
So it's just to me under the first hope neighbours of myself to even for demographics.

584
01:07:02,240 --> 01:07:10,920
And you can do that with reasonable accuracy of what the results are still usually like between 70 and 80 percent accuracy,

585
01:07:10,920 --> 01:07:18,630
even without making the link explicitly to demographics in itself will be interesting to maybe do some clustering on just the type of communication.

586
01:07:18,630 --> 01:07:22,140
Yeah, yeah, yeah, definitely. I think these kind of predictions are based on that.

587
01:07:22,140 --> 01:07:29,940
So, for example, females and males have different communication patterns during the weekdays on the weekends, so that's how they the predictor.

588
01:07:29,940 --> 01:07:37,290
So getting back to your point, if you don't have demographics, maybe you can use that to predict and then use that to condition the matching.

589
01:07:37,290 --> 01:07:47,200
But there is one actual stop in the middle that you don't guarantee that the prediction is perfect.

590
01:07:47,200 --> 01:07:54,200
On the other hand, you can use social media activities to predict. I think Scott hail from the I did something recently.

591
01:07:54,200 --> 01:07:58,210
And they just look at Twitter activities and they use that to predict demographics.

592
01:07:58,210 --> 01:08:02,710
I think it's kind of similar idea.

593
01:08:02,710 --> 01:08:09,180
So I'm wondering in your inference of the Netflix apology, do you have in this case, you had this kind of quadratic function?

594
01:08:09,180 --> 01:08:14,790
Yeah, but I guess it's quite generally two different kinds of functions.

595
01:08:14,790 --> 01:08:24,430
Yeah. So I'm yeah, I'm trying to understand perhaps variations, where have you thought or have you built?

596
01:08:24,430 --> 01:08:29,620
Have you tried to do this with other functions as well to try and see how it, how it looks like?

597
01:08:29,620 --> 01:08:33,340
Or or um. Yeah.

598
01:08:33,340 --> 01:08:38,440
So the answer is that we haven't really tried the other health functions, too.

599
01:08:38,440 --> 01:08:41,950
The thing is that the first we want wanted to be general enough.

600
01:08:41,950 --> 01:08:52,060
Second, the equilibrium needs to be explicitly computable because we want to have this notion of natural Grootboom actions.

601
01:08:52,060 --> 01:08:57,730
If you have, let's say, uh, arbitrary analytical function of the of,

602
01:08:57,730 --> 01:09:03,520
as long as you can take the first or the derivative, you can you fit into the optimisation.

603
01:09:03,520 --> 01:09:10,180
But in that case, there is no economic theory to guarantee that the equilibrium will be stable and unique,

604
01:09:10,180 --> 01:09:16,510
which means you kind of lose something from the, uh, economic interpretation.

605
01:09:16,510 --> 01:09:18,880
Right?

606
01:09:18,880 --> 01:09:26,890
So one of the things that we've discussed a lot in the Summer Institute is is that some of these very detailed types of data that you've talked about,

607
01:09:26,890 --> 01:09:27,520
for example,

608
01:09:27,520 --> 01:09:40,510
mobile networks calling networks or other forms of of rich behavioural data are often in the hands of private companies and hard to get access to.

609
01:09:40,510 --> 01:09:47,020
And while they're very they can shed light to new kinds of ways as you were in the dark,

610
01:09:47,020 --> 01:09:55,240
perhaps you could comment a bit on on on your experiences trying to access these kinds of data, especially, for example, the mobile study in Andorra.

611
01:09:55,240 --> 01:09:59,950
And what were the steps or the pathways that enabled you to get access to them?

612
01:09:59,950 --> 01:10:01,780
Because I think it would be helpful for us here as well.

613
01:10:01,780 --> 01:10:14,830
To know about that Andorra study, I think was based on kind of a contract signed between a few groups in the Media Lab and the government of Andorra,

614
01:10:14,830 --> 01:10:22,450
because at the time they were trying to diversify the economy and then trying to attract more tourists, basically.

615
01:10:22,450 --> 01:10:31,690
And now understand whether they can keep them to stay overnight so that they can, and they have some huge problems in terms of traffic jams.

616
01:10:31,690 --> 01:10:42,970
So so they basically want to work with a few groups in the media lab to build the new models first to solve these problems.

617
01:10:42,970 --> 01:10:52,790
But they also want to make Android like a place of innovation in Europe by testing new ideas like these and self-driving cars.

618
01:10:52,790 --> 01:10:56,770
And that was like five years ago when they were trying to put some vehicles on the

619
01:10:56,770 --> 01:11:03,160
street and then see how people react in that case because of that special relationship.

620
01:11:03,160 --> 01:11:08,000
We got this access to data from the.

621
01:11:08,000 --> 01:11:18,740
Telecom, Andorra, I think, and mostly what I know the way that you got access to this kind of data is through this kind of research contracts.

622
01:11:18,740 --> 01:11:22,280
But I would say it's more and more common compared to 10 years ago, 10 years ago,

623
01:11:22,280 --> 01:11:30,950
it was really very difficult because people do not understand whether it's safe to share it or whether they can actually get some benefit.

624
01:11:30,950 --> 01:11:34,520
But nowadays, I would say specifically for mobile phone data,

625
01:11:34,520 --> 01:11:39,950
it becomes more popular because more and more network providers realise that

626
01:11:39,950 --> 01:11:45,920
there is a reasonably good way to share anonymized data for some benefits.

627
01:11:45,920 --> 01:11:56,750
So, uh, I don't know if you know there is a dedicated conference called Net Mobile, which actually happens the week after next in Oxford.

628
01:11:56,750 --> 01:12:00,890
So that's a dedicated conference for people working on mobile phone data.

629
01:12:00,890 --> 01:12:06,530
So there you can see that they have all sorts of data from different countries.

630
01:12:06,530 --> 01:12:16,970
But in general, I think the problem is that the reason why people do not necessarily want to share the data is because of these ethical concerns.

631
01:12:16,970 --> 01:12:23,780
They are not sure whether it will be safe and what would be the best way to share what is the best way to anonymize.

632
01:12:23,780 --> 01:12:29,390
If they don't know, then it becomes very difficult for them to share a name for your approach, tending to be difficult.

633
01:12:29,390 --> 01:12:35,570
So I think the idea would be to I think people are convinced that they are helpful.

634
01:12:35,570 --> 01:12:39,470
But now this means that the other side of the development needs to catch up,

635
01:12:39,470 --> 01:12:49,220
basically to show that we can have safe and meaningful ways to share the data, which can help us to solve something important problems.

636
01:12:49,220 --> 01:12:54,980
I think that can help us convince the data holder in this case.

637
01:12:54,980 --> 01:12:58,160
But it takes time because I could tell you, if you compare 10 years ago and now,

638
01:12:58,160 --> 01:13:02,720
10 years ago, maybe June, if only a few groups in the world that have this kind of data.

639
01:13:02,720 --> 01:13:11,080
But now I would say much, much more common. Still, I think.

640
01:13:11,080 --> 01:13:15,970
It's kind of a slow process in some ways.

641
01:13:15,970 --> 01:13:29,770
So. I think you have a question regarding the network model that you that you showed before,

642
01:13:29,770 --> 01:13:37,960
when I got that right with the inverse of the policy areas in the model that you get some kind of like the temporal process that I'm influencing,

643
01:13:37,960 --> 01:13:41,680
like my network neighbours and like other neighbours.

644
01:13:41,680 --> 01:13:48,940
And it also comes back to me, doesn't it like that? So in this case, if I understand your question, this is more like a one shot game.

645
01:13:48,940 --> 01:14:01,810
So it's not a sequential game. So it's just like I just observed action where people do not know the decisionmaking, so the other people beforehand.

646
01:14:01,810 --> 01:14:10,540
So they basically make simultaneous actions and the network is basically a network structure in a particular time.

647
01:14:10,540 --> 01:14:23,710
OK, so you observe this simultaneous actions and then you in four network in a given point in time, but is not sequentially done?

648
01:14:23,710 --> 01:14:32,440
But wouldn't I be like if I just solve the the inverse inverse term or I'd be there as well?

649
01:14:32,440 --> 01:14:37,900
Like, uh, like, isn't that wouldn't it look modern?

650
01:14:37,900 --> 01:14:42,210
Kind of like that. My effort depends on my effort somewhat. Yeah, you.

651
01:14:42,210 --> 01:14:48,760
I mean, actually, it's like there is a loop, right? So your your efforts depends on the others and the others are depending on the third person.

652
01:14:48,760 --> 01:14:52,320
It's like that. Isn't it only possible in time like this?

653
01:14:52,320 --> 01:14:57,610
You just model one simultaneously is, uh, not necessary.

654
01:14:57,610 --> 01:15:03,490
You can. You can basically have these examples that I presented at the beginning where basically we say

655
01:15:03,490 --> 01:15:09,220
that say you are in a social network and now you know your neighbours and then in your mind,

656
01:15:09,220 --> 01:15:16,570
you basically work out like what would be the equilibrium, right? If I do something, I can expect, what is the action right?

657
01:15:16,570 --> 01:15:20,260
So if everybody just to if everybody's assumed to be rational,

658
01:15:20,260 --> 01:15:25,570
they will basically compute this equilibrium in their mind and they would basically make a simultaneous action.

659
01:15:25,570 --> 01:15:28,180
And then that's a gripping you observe, basically,

660
01:15:28,180 --> 01:15:36,210
because the idea of recruitment means that if the others do not change, then I don't have the incentive to change.

661
01:15:36,210 --> 01:15:41,210
Right. Yeah. But you can generalise, generalise into to a sequential setting.

662
01:15:41,210 --> 01:15:51,120
I think where you say, OK, I observe the past joint decision of all the people and that may be the network attention the next.

663
01:15:51,120 --> 01:15:58,680
Timestamp. Maybe the actions will depend on the actions in this new point in time.

664
01:15:58,680 --> 01:16:09,910
Plus the game that played in the past, but not here so.

665
01:16:09,910 --> 01:16:17,800
So I have another question about data collection, so one of the I mean, I think like network data are things that, of course,

666
01:16:17,800 --> 01:16:28,180
are very exciting, but also not as as as widely available as as one would like, partly because they're difficult to connect data on.

667
01:16:28,180 --> 01:16:38,500
And I am wondering if based on your examples, given that you do a lot of work on you use with networks,

668
01:16:38,500 --> 01:16:46,120
are there kind of examples of good data sets that are also publicly available that that you know you'd like to talk or share?

669
01:16:46,120 --> 01:16:50,620
Let us know about. Because I mean, this has been recently enough in college.

670
01:16:50,620 --> 01:16:54,430
There was a workshop on, I think, a couple of years ago or last year,

671
01:16:54,430 --> 01:16:58,570
which was on sort of the challenges of network data collection and the fact that there

672
01:16:58,570 --> 01:17:05,230
are just so few available and often there are maybe very small samples if they are.

673
01:17:05,230 --> 01:17:10,920
And so and you mentioned that now a lot more data sources are available, especially for networks.

674
01:17:10,920 --> 01:17:16,240
So perhaps you could just give us more examples of that to leverage those that would be very helpful.

675
01:17:16,240 --> 01:17:20,230
Yeah. If you if you talk about the like or, for example, real world social network data,

676
01:17:20,230 --> 01:17:26,050
at least in the computer science literature, there are many, many repositories.

677
01:17:26,050 --> 01:17:31,330
I would say you can find all sorts of networks in real world situations because

678
01:17:31,330 --> 01:17:35,710
those are the networks that a computer scientists to test the algorithms on.

679
01:17:35,710 --> 01:17:42,340
They can be static or they can be dynamic that can be directed to can be for all different types of applications.

680
01:17:42,340 --> 01:17:52,930
I can share a few links, maybe. And the more difficult thing would be the data that you mentioned first, like the more private.

681
01:17:52,930 --> 01:17:59,650
Yeah. Or data that are privately held by companies and governments.

682
01:17:59,650 --> 01:18:08,290
That's more difficult and that's more difficult also because of the nature of the data, which has a like a higher risk.

683
01:18:08,290 --> 01:18:14,530
If you think about the social network, the risks, it kind of lasts because you just have the Connexions.

684
01:18:14,530 --> 01:18:23,950
If you don't have any other information on the nodes, it's just like a graph. It's a matrix doesn't really allow any identification.

685
01:18:23,950 --> 01:18:36,110
But if you have a trace system, then that's a different story. So.

686
01:18:36,110 --> 01:18:40,880
All right, well, we want to thank Joe Dunn for Joe and for speaking with us.

687
01:18:40,880 --> 01:18:44,660
And if you have any more questions, open the door. Yeah.

688
01:18:44,660 --> 01:18:53,510
OK, so we'll come to lunch if you have more questions. And after lunch is just continue to work on your group, projects will be sending out soon.

689
01:18:53,510 --> 01:18:59,030
A little brief on what we would like in the presentations tomorrow.

690
01:18:59,030 --> 01:19:06,517
So anyone have any questions? OK. Thank you.