Large-Scale Social Network Data Mining with Multi-View Information. Hao Wang

Transcription

1 Large-Scale Social Network Data Mining with Multi-View Information Hao Wang Dept. of Computer Science and Engineering Shanghai Jiao Tong University Supervisor: Wu-Jun Li Hao Wang Multi-View Social Mining 1 / 28

2 Our work 1 Hao Wang, Binyi Chen, Wu-Jun Li. Collaborative Topic Regression with Social Regularization for Tag Recommendation. Proceedings of the Twenty-Third International Joint Conference on Artificial Intelligence (IJCAI), Hao Wang, Wu-Jun Li. Relational Collaborative Topic Regression for Recommendation Systems. IEEE Transactions on Knowledge and Data Engineering (TKDE), (submitted) 3 Hao Wang, Wu-Jun Li. Online Egocentric Models for Citation Networks. Proceedings of the Twenty-Third International Joint Conference on Artificial Intelligence (IJCAI), Hao Wang Multi-View Social Mining 2 / 28

3 Outline 1 Motivation 2 Relational CTR 3 CTR with Social Regularization 4 Online Egocentric Models 5 Conclusion Hao Wang Multi-View Social Mining 3 / 28

4 Outline 1 Motivation 2 Relational CTR 3 CTR with Social Regularization 4 Online Egocentric Models 5 Conclusion Hao Wang Multi-View Social Mining 4 / 28

5 Motivation Hao Wang Multi-View Social Mining 4 / 28

6 Motivation Content: 1 Lee et al., Chen et al., Lipczak et al., 2009 Ratings: 1 Salakhutdinov et al., Herlocker et al., 1999 Hybrid: 1 Purushotham et al., Wang and Blei, Agarwal and Chen, A. Said, 2010 Hao Wang Multi-View Social Mining 5 / 28

7 Contribution 1 Integrate ratings, contents and item networks 2 Significantly improve the accuracy 3 Even less training time 4 Extend to dynamic networks Hao Wang Multi-View Social Mining 6 / 28

8 Outline 1 Motivation 2 Relational CTR 3 CTR with Social Regularization 4 Online Egocentric Models 5 Conclusion Hao Wang Multi-View Social Mining 7 / 28

9 Comparison of article representation Hao Wang Multi-View Social Mining 7 / 28

10 Combination of both Hao Wang Multi-View Social Mining 8 / 28

11 Graphical model of CTR z w k K v v u r u I J Hao Wang Multi-View Social Mining 9 / 28

12 Graphical model of RCTR j N zj, n wj, n sj vj ri, j J e l ' j, j r v u ui ui I k K sj ' vj ' ' ri, j I ' j N zj ', n wj ', n J Hao Wang Multi-View Social Mining 10 / 28

13 Objective function The log-likelihood: L = ρ log σ(η T (s j s j ) + ν) (j,j ) λ r (s j v j ) T (s j v j ) λ e 2 2 η+ T η + λ u 2 + j j u T i u i λ v (v j θ j ) T (v j θ j ) 2 i j log( θ jk β k,wjn ) c ij 2 (r ij u T i v j ) 2. n k i,j Hao Wang Multi-View Social Mining 11 / 28

14 Updating rules For U and V : u i (V C i V T + λ u I K ) 1 V C i R i, v j (UC i U T + λ v I K + λ r I K ) 1 (UC j R j + λ v θ j + λ r s j ), For η + : η +L = ρ (1 σ(η + T π + j,j ))π + j,j λ e η +, l j,j =1 For φ jnk : φ jnk θ jk β k,wjn. For β kw : β kw j φ jnk 1[w jn = w]. n Hao Wang Multi-View Social Mining 12 / 28

15 Datasets Description of datasets citeulike-a citeulike-t #users #items #tags #citations #user-item pairs sparsity 99.78% 99.93% #relations Hao Wang Multi-View Social Mining 13 / 28

16 Experimental results 0.3 CF CTR RCTR Recall M The user-oriented recall of RCTR, CTR, and CF when M ranges from 50 to 300 on dataset citeulike-t. P is set to 1. Similar phenomenon can be observed for other values of P. Hao Wang Multi-View Social Mining 14 / 28

17 Case study Interpretability of the learning latent structures Hao Wang Multi-View Social Mining 15 / 28

18 Outline 1 Motivation 2 Relational CTR 3 CTR with Social Regularization 4 Online Egocentric Models 5 Conclusion Hao Wang Multi-View Social Mining 16 / 28

19 Graphical model of CTR z w k K v v u r u I J Hao Wang Multi-View Social Mining 16 / 28

20 Graphical model of CTR-SR z w k K v v r l A s u r u I J Hao Wang Multi-View Social Mining 17 / 28

21 Objective function The log-likelihood: where L = λ l 2 tr(sl as T ) λ r 2 λ u 2 + j (s j v j ) T (s j v j ) j u T i u i λ v (v j θ j ) T (v j θ j ) 2 i j log( θ jk β k,wjn ) c ij 2 (r ij u T i v j ) 2. n k i,j tr(sl a S T ) = 1 2 = 1 2 J j=1 j =1 J j=1 j =1 J A jj S j S j 2 J [A jj K (S kj S kj ) 2 ] k=1 Hao Wang Multi-View Social Mining 18 / 28

22 Updating rules For U and V : For S: For φ jnk : For β kw : u i (V C i V T + λ u I K ) 1 V C i R i, v j (UC i U T + λ v I K + λ r I K ) 1 (UC j R j + λ v θ j + λ r s j ), S k (t + 1) S k (t) + δ(t)r(t) r(t) λ r V k (λ l L a + λ r I J )S k (t) δ(t) β kw j r(t) T r(t) r(t) T (λ l L a + λ r I J )r(t) φ jnk θ jk β k,wjn. φ jnk 1[w jn = w]. n Hao Wang Multi-View Social Mining 19 / 28

23 Experimental results SCF SCF+LDA CF TAGCO CTR CTR SR 0.2 Recall P Recall@50 for all methods in citeulike-a. Hao Wang Multi-View Social Mining 20 / 28

24 Case study Example articles with recommmended tags Hao Wang Multi-View Social Mining 21 / 28

25 Outline 1 Motivation 2 Relational CTR 3 CTR with Social Regularization 4 Online Egocentric Models 5 Conclusion Hao Wang Multi-View Social Mining 22 / 28

26 From DEM to OEM Three parts of variables: 1 Model parameters β 2 Link features 3 Topic features Dynamic Egocentric Models (DEM, adapting only Model parameters): m L(β) = e=1 i=1 exp(β T s ie (t e )) n Y i (t e ) exp(β T s i (t e )) Online Egocentric Models (OEM, adapting all three parts): minimize log L(β, ω) + λ n ω k θ k 2 2 k=1 subject to : ω k 0, 1 T ω k = 1, Hao Wang Multi-View Social Mining 22 / 28

27 Updating rules Using coordinate descent: 1 Online β Step: L w (β) = 2 Online Topic Step: f ω k = + x+q 1 e=x+q W t i=1 p a i + i=1 q u=p+1 exp(β T s ie (t e )) n. Y i (t e ) exp(β T s i (t e )) p i=1 + 2λ(ω k θ k ) a i α i exp(a T i ω k ) A i + α i exp(a T i ω k) b u γ u exp(b T u ω k ) B u + γ u exp(b T u ω k ) Hao Wang Multi-View Social Mining 23 / 28

28 Datasets Information of data sets Data Set #Papers #Citations #Unique Times arxiv-th arxiv-ph Data set partition for building, training and testing. Data Sets Building Training Testing arxiv-th arxiv-ph Hao Wang Multi-View Social Mining 24 / 28

29 Experimental results 8 8 Average log likelihood DEM OEM beta OEM appr OEM full Paper batches Average log likelihood DEM OEM beta OEM appr OEM full Paper batches (a) arxiv-th (b) arxiv-ph Recall DEM 0.3 OEM beta OEM appr OEM full Paper batches (c) arxiv-th(k=250) Recall DEM 0.25 OEM beta OEM appr OEM full Paper batches (d) arxiv-ph(k=250) Hao Wang Multi-View Social Mining 25 / 28

30 Outline 1 Motivation 2 Relational CTR 3 CTR with Social Regularization 4 Online Egocentric Models 5 Conclusion Hao Wang Multi-View Social Mining 26 / 28

31 Conclusion Data Mining with Multi-View Information: 1 Relational CTR: Social information as prior 2 CTR with Social regularization: Social information as observation 3 Online Egocentric Models: Dynamic social information Hao Wang Multi-View Social Mining 26 / 28

32 Publication 1 Hao Wang, Binyi Chen, Wu-Jun Li. Collaborative Topic Regression with Social Regularization for Tag Recommendation. Proceedings of the Twenty-Third International Joint Conference on Artificial Intelligence (IJCAI), Hao Wang, Wu-Jun Li. Online Egocentric Models for Citation Networks. Proceedings of the Twenty-Third International Joint Conference on Artificial Intelligence (IJCAI), Hao Wang, Wu-Jun Li. Relational Collaborative Topic Regression for Recommendation Systems. IEEE Transactions on Knowledge and Data Engineering (TKDE), (submitted) Hao Wang Multi-View Social Mining 27 / 28

33 Q & A Hao Wang Multi-View Social Mining 28 / 28