Who Should Share What? Item-level Social Ifluence Prediction for Users and Posts Ranking RunquanXie E-mail: RunquanXie@126.com 厦门大学计算机科学系 2011年11月 2015/4/13 Author information Peng Cui Research Assistant Shaowei Liu Ph.D. Shiqiang Yang Professor 2015/4/13 ABSTRACT 1.Two core dimensions in a social network:information people basic behavior:people sharing information spread 2. Who Should Share What?(predicting item-level social influenece) Two information retrieval scenarios (1) Users ranking an post (2) Posts ranking an user 3. Formulate this problem : the estimation of a user-post matriX, each entry : the strength of influence of a user given a web post. 4. Hybrid Factor Non-Negative Matrix Factorization approach Devise an efficient projected gradient method 2015/4/13 INTRODUCTION •With the rapid proliferation : Facebook, Renren etc., more and more user profiles,interactions, and collective intelligence ( tags, comments, etc.) are available online. •a new perspective for information retrieval applications: • more focus should put on user collaborative information. •MeanWhile, new search scenarios: •web people search and relationship search emerged. •challenge to traditional information retrieval: • how to effectively handle the social information? 2015/4/13 INTRODUCTION 1.A key concept : Social influence a prevalent [普遍]and complex force governing the dynamics of social network. 2.a clear need for techniques to analyze social influence and this issue in information retrieval field has still not been well studied. 3. The existing social influence analysis research[summarized into a diagram] Who(A) influences Whom(B) given What(C)? A:a single user B:the whole network or a single user or the A’s friends (neighborhoods) in different previous works 2015/4/13 INTRODUCTION C : three lines of research 1.Structure-level [C is empty] 2.Topic-level [C is a topic] 3.Item-level [C is an item (such as a web page, product etc.)] rare in previous research. 2015/4/13 Motivating Application Definition on Wiki : social influence occurs when an individual’s thoughts, feelings or actions are affected by other people. In Renren, when a user share a web post, a portion of her friends will click, comment, or even forward the post, which are three levels of influence . In this paper, we only consider click action. That is, the social influence of a user on her friends given a web post is defined as the number of her friends who click the shared web post. 2015/4/13 Challenges 1.User-post specific User-post matrix 2. Sparsity effective prior knowledge for user and post grouping to alleviate the sparsity problem. User-user similarity matrix and post topic distribution matrix 3.Complex factors. For example, the total number of friends, the tie strength between the user and her friends, the semantics of web posts, etc., How to select the effective factors and integrate these complex factors in one predictive model is also one of the focus of our work. 2015/4/13 Contributions 1. Formulates the item-level social influence prediction problem formally with HF-NMF, and devise an efficient projected gradient method to solve it. 2. Support the applications such as influencer ranking and information recommendation by user-post matrix ranking in two directions. 3. social influence extendable to the influence on all the friends and the friends of friends etc. 4. We conducted intensive experiments on real social network datasets, and the results show that the HF-NMF can achieve a better performance compared with other competitors. 2015/4/13 PRELIMINARIES : Necessity of Item-level Social Influence 1.The dataset http://renren.com/, 150 million active users. 2.A user can generate a post or share a web page and as the number of friends is different for each user, the upper bound of users’ social influence strengths are also different. In order to make the strength of influence be measured in a unified scale for different users for the sake of observational and modeling study, we use the proportion of friends who click the shared post as the measure. 3.Figure 1 2015/4/13 Figure1 2015/4/13 Figure1 (1) different users have different influence power to their friends (2) different posts have different influence power (3) users’ influences differently for different posts Conclusion : Only item-level social influence can reveal the users’ real influence on friends, and the strength of influence should definitely be user-post specific. 2015/4/13 Predictive Factors 1. User-specific factors. Although users’ social influence vary with the shared posts, the average of the social influences over posts determines the overall social influence of a user. We regard the factors that affect users’ overall social influence (excluding the posts) as he user-specific factors. 2. Post-specific factors. posts’ overall social influence 3. User-post specific factors. The social influence of a user given a post cannot be well approximates only by the user and post-specific factors. The factors indicating the interactions between users and posts. 2015/4/13 Predictive Factors emphasizing 1.The user and post factors are essential for the predictive modeling. 2. The user-post interactions are very sparse. alleviate : find effective factors to “group” those users and posts 3.The user and post-specific factors also provide some effective prior knowledge to complement the inference from pure user-post interactions. 4. How to find out the effective predictive factors?? Finally : Two user-oriented factors: the percentage of active friends, the average social tie strength between a user and her friends, One post-specific factor: the topic distribution of a post’s content. 2015/4/13 demonstrate the validity of selected factors Randomly select 10 users. Define user social influence : the average percentage of her friends who click the shared post over all shared posts. 2015/4/13 Post topic distribution Hypothesis : the posts with similar contents ( similar topic distributions) often induce similar social influences. Validate: we randomly select ten groups of web posts, and posts in the same group have similar topic distributions. Define web post social influence same with above. Then calculate and plot Conclusion : 1.most variances inside topic groups are smaller than that across groups 2.The introduction of topical grouping :reduce down the uncertainty of social influence. 2015/4/13 ALGORITHM 1.Suppose :M users ui and N postings pj. N(ui) : ui ’s first-order friends. 2. Item-level social influence the strength of ui’s influence on N(ui) : fij, is the number of ui’s friends whO clicked pj . 3. Social influence prediction[M*N] The social influence prediction is to predict the unobserved social influences fˆij based on the observed fij ’s and those predictive factors. 2015/4/13 Formally X ∈ RM×N Gi : the number of ui’s friends different users have different numbers of friends Problem: predicting the unobserved entries in X. 2015/4/13 Hybrid Factor Non-negative Matrix Factorization • Suppose:a joint latent space for both users and posts with dimensionality k • Ui : an user vector ui ∈ Rk • Pj : a post vector vj ∈ Rk 2015/4/13 (HF-NMF) Focus more on the valid[有效] elements Sharing matrix : The severe sparsity of X?? 2015/4/13 Incorporating User-Specific Factors 1.The percentage of active friends: 2. the average friend tie strength: 2015/4/13 1.the similarity between ui and uj 2. the user-user similarity matrix 2015/4/13 Incorporating Post-Specific Factors 1.The post content matrix social influence is strongly correlated with the content of the web posts 2. Finally 2015/4/13 Solution 1.a nonnegativity projection operator P[ ] formatrix A Problem: Variants of projected gradient methods differ on selecting the step size 2015/4/13 Algorithm 1 Projected Gradient Ensures the sufficient decrease of the function value per iteration. • By trying the step sizes 2015/4/13 Algorithm 2 Improved Projected Gratient 1. Searching is the most time consuming operation 2. .The trick: and guess either increases or decreases 2015/4/13 Algorithm 2 Improved Projected Gratient • a better initial guess of at each iteration and allows to be larger than one. 2015/4/13 2015/4/13 交替下降 => for fix V G update U fix U G update V fix U V update G end for 2015/4/13 EXPERIMENTS 1.DataSet [ at least 6 days before the crawling date ] 2015/4/13 Comparative Methods 1. 2. 3. 4. Logistic Regression Cox Proportional Hazards Regression User Averaging Influence Post Averaging Influence 5. Basic Non-NegativeMatrix Factorization 2015/4/13 Comparative Methods 6.User Factors Constrained NMF 7. Post Factors Constrained NMF 2015/4/13 Evaluation Measures 1. 2. T-measure R-measure 1 : perfect ranking 2015/4/13 Parameter Settings : Tradeoff Parameters α =0.0001, β = γ = δ = 0.01 2015/4/13 Parameter Settings : Dimensionality of the Hidden Space 2015/4/13 Number of Iterations 2015/4/13 Prediction Performance 2015/4/13 User and Post Ranking Performance 2015/4/13 Comparison with Other Methods 2015/4/13 CONCLUSION 1.Hybrid Factor Non-Negative Matrix Factorization method to incorporate these predictive factors for user-post specific social influence prediction 2. devise an efficient Projected Gradient method for HFNMF solution. 2015/4/13 Department of Computer Science, Xiamen University, Nov, 2011 2015/4/13