Social Prediction in Mobile Networks: Can we infer users emotions and social ties? Jie Tang Tsinghua University, China 1 Collaborate with John Hopcroft, Jon Kleinberg (Cornell) Jinghai Rao (Nokia), Jimeng Sun (IBM TJ Watson) Tiancheng Lou, Wenbin Tang, Honglei Zhuang, Yuan Zhang (Tsinghua)
Motivation Social behavior VS. Emotion change 2
Motivation Emotion Social stimulates behavior the mind 3000 Emotion times change quicker than rational VS. thought!!! It's an emotional world we live in! Six degree vs. Three degree [Nature; BMJ] 3
Motivation: A Happy System Can we predict users activities and emotion? 4
Motivation: Inferring Social Ties From Home 08:40 From Office 11:35 From Office 15:20 Both in office 08:00 18:00 From Office 17:55 0.89 0.77 0.98 0.63 0.70 Friends Other From Outside 21:30 0.86 5
6 Motivation: RideSharing
MoodCast: Emotion Prediction via Dynamic Continuous Factor Graph Model ICDM 10, IEEE Trans. on Affective Computing 11 7
Happy System 8 Can we predict users emotion?
荷 塘? Dorm? Observations 教 室??? GYM Activity correlation Location correlation (Red-happy) KO 9
Observations (cont.) (a) Social correlation Social correlation (a) Implicit groups by emotions 10 Temporal correlation
Observations (cont.) We should not split the data into different time windows Calling (SMS) correlation 11
MoodCast: Dynamic Continuous Factor Graph Model MoodCast Social correlation g(.) Jennifer Allen Mike Temporal correlation h(.) Jennifer yesterday Neutral Allen Happy Jennifer today Happy Neutral Mike Predict Jennifer tomorrow? Attributes f(.) location call sms Our solution 1. We directly define continuous feature function; 2. Use Metropolis-Hasting algorithm to learn the factor graph model. 12
Problem Formulation Time t G t =(V, E t, X t, Y t ) Emotion: Sad Time t-1, t-2 Attributes: - Location: Lab - Activity: Working Learning Task: 13
Dynamic Continuous Factor Graph Model Time t Time t : Binary function 14
Model Learning y 5 y 4 y ' 3 y 3 y 2 y 1 Attribute Social Temporal 15
16 MH-based Learning algorithm
Experiment Data Set Baseline SVM SVM with network features Naïve Bayes Naïve Bayes with network features Evaluation Measure: Precision, Recall, F1-Measure #Users Avg. Links #Labels Other MSN 30 3.2 9,869 >36,000hr LiveJournal 469,707 49.6 2,665,166 17
18 Performance Result
Factor Contributions Mobile All factors are important for predicting user emotions 19
Inferring Social Ties in Mobile Networks PKDD 2011 (Best Paper Runnerup), WSDM 2012 20
Real social networks are complex... Nobody exists only in one social network. Public network vs. private network Business network vs. family network However, existing networks (e.g., Facebook and Twitter) are trying to lump everyone into one big network FB tries to solve this problem via lists/groups However Google+ which circle? Users do not take time to create it. 21
Even complex than we imaged! Only 16% of mobile phone users in Europe have created custom contact groups users do not take the time to create it users do not know how to circle their friends The fact is that our social network is black- 22
Problem Formulation Input: G=(V,E L,E U,R L,W) Partially Other Labeled Network?? Other Friend? V: Set of Users E L,R L : Labeled relationships E U : Unlabeled relationships 23 Input: G=(V,E L,E U,R L,W) Output: f: G R
Basic Idea V 1 V 3?? Friend V 2 User Node?? r 24 r 56 Other r 45 Relationship Node 24
Partially Labeled Pairwise Factor Graph Model (PLP-FGM) Constraint factor h 25 Input: Social Network v 2 Problem: v 4 v 3 v 5 v 1 PLP-FGM y 12 =Friend y 12 =advisor y 12 Latent Variable h (y 12, y 21 ) g (y 12, y 34 ) g (y 12,y 45 ) f(x 2,x 1,y 21 ) f(x 1,x 2,y 12 ) y y 21 =Friend =advisee 21 y 34 =? y 34 f(x 3,x 4,y 34 ) r 12 r 34 r 21 g (y 45, y 34 ) y 45 y 34 y 16 =coauthor y 16 =Other f(x 4,x 5,y 45 ) r 34 r 45 y 34 =? f(x 3,x 4,y 34 ) Correlation factor g relationships Attribute factors f For each Input relationship, identify which type Model Map has relationship the highest probability? to nodes in model Example: Call A makes frequency call to between B immediately two users? after the call to C. Partially Labeled Model
Solutions (con t) Different ways to instantiate factors We use exponential-linear functions Attribute Factor: Correlation / Constraint Factor: Log-Likelihood of labeled Data: 26
Learning Algorithm Maximize the log-likelihood of labeled relationships Expectation Computing Loopy Belief Propagation Gradient Ascent Method 27
Still Challenges? Questions: - How to obtain sufficiently training data? - Can we leverage knowledge from other network? 28
Inferring Social Ties Across Networks Input: Heterogeneous Networks Reviewer network Adam review Output: Inferred social ties in different networks Adam Bob review Product 1 Bob distrust distrust trust review Chris trust Chris Danny review Communication network Product 2 Knowledge Transfer for Inferring Social Ties Danny From Home 08:40 Both in office 08:00 18:00 Family Colleague From Office 11:35 Colleague From Office 15:20 From Outside 21:30 From Office 17:55 What is the knowledge to transfer? Friend Colleague Friend 29
Social balance theory Structural hole theory Social Theories friend A friend friend A non-friend friend A friend non-friend A non-friend B friend C B non-friend C B non-friend C B non-friend (A) (B) (C) (D) C 30
Social Theories Structural hole Social balance theory Structural hole theory Structural hole 31
Transfer Factor Graph Model 32 Coauthor network mobile y y 2 =? 4 =? y 2 y 4 h (y 3, y 4, y 5 ) TrFG model y 5 y y 1 1 =1 y 5 =1 Input: social network y 3 h (y 1, y 2, y 3 ) y 3 =0 y 6 y 6 =? f (s v 3, s 3,y 3 ) 3 5 v 6 f (u 5,s 5, y 5 ) 4 f (s 1, u 2,y 1 ) f (u 2, s 2,y 2 ) v f (u 3 4, s 4,y 4 ) 4 6 u 2, s 2 f (s 6, u 6,y 6 ) 2 v 5 (v 2, v 3 ) u 4, s 4 u v 5, s 5 2 u 1, s 1 u 3, s 3 (v 4, v 5 ) (v u 6, s 6 (v 4, v 6 ) 1 v 2, v 1 ) 1 (v 4, v 3 ) (v 6, v 5 ) Observations y y 2 =? 4 =? y 2 y 4 h (y 3, y 4, y 5 ) TrFG model y 5 y y 1 1 =1 y 5 =1 Input: social network y 3 h (y 1, y 2, y 3 ) y 3 =0 y 6 y 6 =? f (s v 3, s 3,y 3 ) 3 5 v 6 f (u 5,s 5, y 5 ) 4 f (s 1, u 2,y 1 ) f (u 2, s 2,y 2 ) v f (u 3 4, s 4,y 4 ) 4 6 u 2, s 2 f (s 6, u 6,y 6 ) 2 v 5 (v 2, v 3 ) u 4, s 4 u v 5, s 5 2 u 1, s 1 u 3, s 3 (v 4, v 5 ) (v u 6, s 6 (v 4, v 6 ) 1 v 2, v 1 ) 1 (v 4, v 3 ) (v 6, v 5 ) Observations Triad-based factor Bridge via social theories
Mathematical Formulation Features defined in different networks Triad-based features shared across networks 33
Data Sets Epinions a network of product reviewers: 131,828 nodes (users) and 841,372 edges trust relationships between users Slashdot: 82,144 users and 59,202 edges friend relationships between users Mobile: 107 mobile users and 5,436 edges to infer friendships between users 34
Results Data Set Method Prec. Rec. F1 Mobile Epinions to Mobile (40%) Slashdot to Mobile (40%) SVM 89.83 59.55 71.62 CRF 94.55 54.17 68.87 PFG 100.00 59.24 74.40 TranFG 82.39 83.44 82.91 TranFG 72.58 85.99 78.72 SVM and CRF are two baseline methods; PFG is the proposed partially-labeled factor graph model; TranFG is the proposed transfer based factor graph model. 35
Varying the percent of the labeled data Epinions-to-Mobile Slashdot-to-Mobile 36
Factor contribution analysis SH-Structural hole; SB-Social balance. 37
38 Conclusions Moodcast: emotion prediction Emotion stimulates the mind 3000 times quicker than rational though; We demonstrate that it is possible to accurately predict users emotions in mobile network. Inferring social ties different types of social ties have essentially different influence on people; By incorporating social theories, our proposed model can significantly improve (+4-14%) the inferring accuracy.
Emotion: Future Work Emotion diffusion in the mobile network; Predicting activities and emotions simultaneously. Inferring social ties: Inferring complex relationships between users, e.g., family, colleague, manager-subordinate; Active learning for inferring social ties. 39
Related Publications Jie Tang, Tiancheng Lou, and Jon Kleinberg. Inferring Social Ties across Heterogenous Networks. WSDM 12. Chi Wang, Jiawei Han, Yuntao Jia, Duo Zhang, Yintao Yu, Jie Tang, Jingyi Guo. Mining Advisor-Advisee Relationships from Research Publication Networks. KDD 10. Wenbin Tang, Honglei Zhuang, and Jie Tang. Learning to Infer Social Relationships in Large Networks. PKDD'11. (Best Student Paper Runner-up) Jie Tang, Yuan Zhang, Jimeng Sun, Jinghai Rao, Wenjing Yu, Yiran Chen, and ACM Fong. Quantitative Study of Individual Emotional States in Social Networks. IEEE Transactions on Affective Computing. 2011 Yuan Zhang, Jie Tang, Jimeng Sun, Yiran Chen, and Jinghai Rao. MoodCast: Emotion Prediction via Dynamic Continuous Factor Graph Model. ICDM'10. Chenhao Tan, Jie Tang, Jimeng Sun, Quan Lin, and Fengjiao Wang. Social Action Tracking via Noise Tolerant Time-varying Factor Graphs. KDD 10. Jie Tang, Jimeng Sun, Chi Wang, and Zi Yang. Social Influence Analysis in Largescale Networks. KDD'09. 40
Thanks! HP: http://keg.cs.tsinghua.edu.cn/jietang/ System: http://arnetminer.org 41