Automatic mapping and modeling of human networks

ALEX (SANDY) PENTLAND

THE MEDIA LABORATORY CAMBRIDGE

PHYSIC A: STATISTICAL MECHANICS AND ITS APPLICATIONS

2007

1. Introduction2. Socioscopes3. Reality mining4. Social signals5. Practical concerns6. ConclusionsComments

Outline2

Studies on office interactions [1] 35–80% of work time in conversation, 14–93% of work time in communication 7–82% of work time in meetings

The properties of human networks : Location context: work, home, etc. Social context: with friends, co-workers, boss, family, etc. Social interaction: are you displaying interest, boredom

etc.

To obtain solid, dynamic estimates of the users’ group membership and the character of their social relationships.

1. Introduction (1/2)3

[1] T. Allen, Architecture and Communication Among Product Development Engineers, MIT Press, Cambridge, MA, 1997, pp. 1–35.

Using this data to model individual behavior as a stochastic process allows prediction of future activity.

The key to automatic inference of information network parameters is the recognition

Standard methods, surveys subjectivity and memory effects, out-of-date. Even information is available, need to validate or

correct by automatic method we present statistical learning methods

wearable sensor data to estimates user’s interaction

1. Introduction (2/2)4

mapping and modeling human networks the conceptual framework used in biological

observation, such as apes in natural surroundings

natural experiments such as twin studies,

but replacing expensive and unreliable human observations with automated, computer-mediated observations. accurately and continuously track the behavior recording with near perfect accuracy.

2. Socioscopes5

Using mobile telephones, electronic badges, and PDAs Tracking the behavior of 94 people in two

divisions of MIT the business school and the Media Laboratory

between 23 and 39 years of age the business school students a decade older than

the Media Lab students. 2/3 male and 1/3 female half were raised in America.

Imaginary Socioscope6

The first part: ‘smart’ phones to observe gross behavior (location, proximity)

continuously over months 330,000 h of data , the behavior of 94 people, 35

yearsThe second part: electronic badges

record the location, audio, and upper body movement to observe for fine-grained behavior (location,

proximity, body motion) over one-day periods The third part: a microphone and software

to analyze vocalization statistics with an accuracy of tenths of seconds

Three main parts of the Socioscope

7

Four main types of analysis: characterization of individual and group

distribution and variability using an Eigenvector or principal components analysis

conditional probability relationships between individual behaviors known as ‘influence modeling’

accuracy with which behavior can be predicted with equal types I and II error rates

comparison of these behavioral measures to standard human network parameters.

2. Socioscopes (4/5)8

Eigenbehaviors provide an efficient method for learning and classifying user behavior [9].

Given behaviors Γ1, Γ2, . . . ,Γm for a group of M people, the average behavior of the group can be defined

by

To deviate an individual’s behavior from the mean. A set of M vectors, Φ = Γi - Ψ,

3. Reality mining9

[9] N. Eagle, A. Pentland, Eigenbehaviors: Identifying Structure in Routine, October 2005, see TR 601 hhttp://hd.media.mit.edui.

Γi(x,y), 2-D location information a low-dimensional

‘behavior space’, spanned by their

Eigenbehaviors

Fig. 110

Principle Components Analysis, PCA a set M orthonormal vectors, un, which best

describes the distribution of the set of behavior data when linearly combined with their respective scalar values, λ n.

Covariance matrix of Φ

Where

The Eigenbehaviors can be ranked by the total amount of variance in the data for which they account, the largest associated Eigenvalues.

3.1. Eigenbehavior modeling11

The main daily pattern, observed subjects leaving their sleeping place to spend

time in a small set of locations during the daylight hours

breaking into small clusters to move to one of a few other buildings during the early night hours and weekends

then back to their sleeping place. Over 85% of the variance in the behavior of

low entropy subjects can be accounted for by the mean vector alone.

3.2. Human Eigenbehaviors (1/2)12

the top three Eigen behavior components the weekend pattern, the working late pattern, and the socializing pattern.

The ability to accurately characterize peoples’ behavior with a low-dimensional model means automatically classify the users’ location context the system to request that the user label locations can achieve very high accuracies with limited

user input.

3.2. Human Eigenbehaviors (2/2)13

Behavioral structure Conditional probability to predict the behavior Two main sub-networks

during the day in the evening

Critical requirement for automatic mapping and modeling of human networks to learn and categorize user behavior accurately capture the dynamics of the network.

3.3. Learning influence (1/2)14

Coupled Hidden Markov Models, CHMMs [10-12] to describe interactions between two people the interaction parameters

limited to the inner products of the individual Markov chains.

The graphical model for influence model

behavior has the same first-order Eigen structure

it possible to analyze global behavior

3.3. Learning influence (2/2)15

[10] A. Pentland, T. Choudhury, N. Eagle, S. Push, Human Dynamics: Computation for Organizations, Pattern Recognition, vol. 26, 2005, pp. 503–511, see TR 589 hhttp://hd.media.mit.edui.[11] W. Dong, A. Pentland, Multi-sensor data fusion using the influence model, IEEE Body Sensor Networks, April, Boston, MA, 2006, see TR 597 hhttp://hd.media.mit.edui.[12] C. Asavathiratham, The influence model: a tractable representation for the dynamics of networked Markov chains, in: Department of EECS, 2000, MIT, Cambridge.

Using the influence model to analyze the proximity data from our cell phone experiment

we find that Clustering the daytime influence relationships 96% accuracy at identifying workgroup affiliation 92% accuracy at identifying self-reported ‘close’

friendships.Similar findings, using the badge platform.

the combination of influence and proximity predicted whether or not two people were affiliated with the same company with 93% accuracy [6].

3.4. Influence modeling16

People are able to ‘size up’ other people from a very short period of observation [13, 14]. linguistic information from observation, to accurately judge prospects for friendship, work

relationship, negotiation, marital prospects

we developed methods for automatically measuring some of the more important types of social signaling [7]. Excitement, freeze, determined and

accommodating.

4. Social signals17

Can predict human behavior? without listening to words or knowing about the

people involved. By linear combinations of social signal

features to accurately predict human behaviors. who would exchange business cards at a

meeting; which couples would exchange phone numbers

at a bar; who would come out ahead in a negotiation; who was a connector within their work group;

18

Predict human behavior

Continuous analysis interactions within an organization may seem reasonable and if misused, could be potentially dangerous.

Conversation postings: the data should be shared, private, or

permanently deleted. Decided by individuals.

Demanding environments: the environmental demands may supersede

privacy concerns.

5. Practical concerns19

human behavior is predictable than is generally thought, and especially predictable from others.

This suggests that humans are best thought of social intelligences

rather than independent actors.As a consequence

can analyze behavior using statistical learning tools such as Eigenvector analysis and influence modeling,

to infer social relationships without to understand the detailed linguistic or cognitive structures surrounding social interactions.

6. Conclusions20

經由 human network 找出人與人間的關係及其建立 model 的作法

在我們的運用是找出 criminal 及找 criminal 的同伙

瞭解將行動電話及不同的 sensor 等如何運用在human network

Human network 對 prediction 的幫助為何 ?

Comments 21