User Modeling In Social Cooperative Behaviors And Its Applications

Recently, with the rapid development of social network and the highly integra-tion of social factors with novel business modes, social network services (SNS) has grown at a breathtaking rate over the past few years. These services not only satisfy users’requirement on social interactions, but also urge the revolution of SNS and raise the abundant social cooperative behaviors. On one hand, the traditional social networks based on familiars in real world are now extended as " anonymous " connec-tions according to their preferences; on the other hand, the social-based P2P diffusion channel further enhance the traditional "broadcast" spreading from authorized infor-mation sources.Therefore, deep analysis on these social cooperative behavior records will strongly support the business intelligence services, which is of great value in re-search and business. Based on the above, in this dissertation, we introduce several recent research efforts on user modeling and applications in social cooperative behav-iors analysis. Specially, by mining the cooperative behavior and interaction (diffusion) records on social network services, we first propose three research directions for social business intelligence, namely user preference modeling, user decision-making analysis, and posterior effects of social learning schemes. Furthermore, from each perspective of above directions, we propose an approach for users’ common interests modeling based on topic-sensitive interactions, with the application of social media annotations; an approach for users’decision analysis based on dynamic social influence, with the ap-plication of social event attendance prediction; an approach for social learning scheme based on latent social connections mining, with the application of future behavior pat-tern estimation. Then a complete framework combines both theoretic and application research could now be formed. To be specific, the research contribution of this disser-tation can be summarized as follows.First, we propose an approach for mining topic-sensitive interaction/diffusion be-haviors with users to reveal users’preference in different social scenarios, then a novel user description techniques could be designed to further support the automatic annota-tions task of social multimedia files. Traditionally, prior arts mainly focus on analyzing individual records, and social interactions could be completely ignored. Therefore, in this section, we define the novel " common-interest based social diffusion " (CIDM), and then formulate the historical diffusion record as the maximal likelihood due to combined effects of users’common interests and information topics. According to the definition, a novel two-stage framework is designed. With the optimal "reproduction" of diffusion records, in the training stage, the common interests could be revealed, and then in the test stage, media contents could be automatically labeled with greedy based algorithms. Extensive experiments on a real-world data set show that our approach outperforms the baselines in the social media annotations task, which validates the potential of social diffusion analysis in user preference modeling. Finally, to ease the computational bur-den of our approach, we further design a ranking based algorithm to rapidly capture dense subgraph in diffusion records streams, which achieves similar performance with much less time spending, then the application of our approach could be further ensured.Second, we propose an approach for simulating the dynamic social interactions and social influences on users within the decision-making process, and further support the application scenario of social event participation predication. Traditionally, prior arts mainly omit the effects of social connections, or simply rely on the static social features or constraints in optimization tasks, thus the social network structure, as well as the so-cial influence factors have been largely ignored. Therefore, in this section, we define the novel "dynamic social influence" within the decision-making process to reflect the mutual dependence of potential attenders. According to the definition, we merge the so-cial influence factors to the formulation of thresholds, while the tendency formulation is based on the users’preference as well as the tolerance of cost, then the correspond-ing two-stage framework could be summarized to combine these complicated factors. Finally, extensive experiments on a real-world data set show that our approach could better predict users’decision in social event participation, which validates the potential of social influence analysis in decision-making analysis.At last, we propose an approach for mining the latent social learning scheme with users, which further supports the prediction of users’future behavior pattern. A crit-ical challenge along this line is that traditional social diffusion modeling could only describe the one-time diffusion process without posterior effects, and the long-term in-fluence could hardly be summarized. Therefore, in this section, we adopt the social diffusion models with integrating the users’behavior modeling, and then owing the fluctuation of behavior pattern ratio to the partial order of social influence. According to the definition, the corresponding two-stage framework could be designed, to reveal the latent social connections and strength of social learning with optimizing the ranking of behavior pattern. After that, prediction of future behavior patterns could be achieved. Finally, extensive experiments on a real-world data set show that our approach could better understand users’future behaviors without extra information, which validates the potential of social learning scheme in user behavior modeling.