Like like alike Joint friendship and interest propagation

Like like alike – Joint friendship and interest propagation in social networks Shuang Hong Yang Georgia Tech Bo Long Yahoo! Labs Nara Sadagopan, Zhaohui Zheng Yahoo! Labs Aug. 11 th, 2010 Alex Smola Yahoo! Research Hongyuan Zha Georgia Tech

Outline • Background • Joint Friendship-Interest Propagation • Experiments • Discussion 2

Background • Social network = friend network + interest network users connect to their friends users interact with service items (applications, ads, games, movies, . . . ) 3

Two key tasks • Friendship Propagation connecting people to real friends – boost traffic & user population, make the social graph denser. . . • Interest Propagation targeting services to people interested – boost revenue, increase user participation, make the interest graph denser. . . These two tasks are usually addressed separately with different methodologies. 4

Homophily: • The social effect: – People connected to each other tend to have similar interest; – People with similar interest are more likely to be friends. • Hints: Freindship and interest evidences are – highly correlated (Y! pulse: higher interest-correlation between connected users) – mutually reinforcing if modeled jointly Friendship and interest should be propagated jointly! 5

Outline • Background • Joint Friendship-Interest Propagation • Experiments • Discussion 6

Friendship-Interest Propagation (FIP) Exploit Homophily to established an integrated network for joint propagation of friendship and interest. 7

The FIP Model • Modeling interests: collaborative filtering i: user j: item y: interest indication φ: latent profiles xi: user features (age, gender, income) xj: item features (words, visual features) 8

The FIP Model • Modeling friendship: latent-factor-based random walk i, i': user s: friendship connection φ: latent profiles xi: user features (age, gender, income) 9

The FIP Model i: user j: item y: interest indication s: friendship connection φ: latent profiles xi: user features (age, gender, income) xj: item features (words, visual features) 10

The FIP model • Model specification: 11

Connection with other models • VS Collaborative filtering – FIP induces a diffusion kernel for users – From CF FIP = from Euclidean Riemannian space = Gaussian kernel information diffusion kernel • VS Random walk – FIP uses latent factor based random walk – avoids direct manipulation of large matrix, hence more scalable 12

Optimization • Overall objective: Dyadic factorization Content factorization Regularization 13

Optimization • Loss functions • Regularizer – L 2, L 1, Ky-Fan, etc 14

Optimization • Stochastic gradient descent in parallel • Feature Hashing to reduce memory overload 15

Bias Correction • Bias Correction – Observations (for both interest and friendship) are sparse with exclusively positive interactions – Absence of negatives leads to inevitable overfitting, e. g. , all the incoming dyadic interactions are predicted positive – Selection bias correction: treat missing observations as very-weak negative observations: 16

Outline • Background • Joint Friendship-Interest Propagation • Experiments • Discussion 17

Experiments • Data – – A subset of Yahoo! Pulse data. 1. 2 M users, 386 items 6. 1 M friend connections 29 M interest indications 18

Experiments • Interest propagation [in terms of service recommendation] 19

Experiments • Interest propagation [in terms of service recommendation] 20

Experiments • Friendship Propagation [in terms of friend suggestion] 21

Experiments • Friendship Propagation [in terms of friend suggestion] 22

Experiments • Bias correction 23

Outline • Background • Joint Friendship-Interest Propagation • Experiments • Discussion 24

Summary and discussion • FIP – Friendship and interest in social networks is correlated and mutually reinforcing – Much higher performance could be achieved by coupling interest and friendship modeling – FIP bridges two different methodologies: collaborative filtering and random walk. 25

Thanks! Any comments would be appreciated! 26