中央研究院 資訊科學研究所

The link prediction problem can be intuitively described as “given a snapshot of a social network, can we infer which new interactions among its members are likely to occur in the near future?” In social network analysis, relationship prediction among people in the interpersonal network is a broadly discussed problem. Nevertheless, when modeling a real network as a heterogeneous information network instead of a homogeneous one, this problem becomes more challenging. In this talk, we address two example applications, i.e., predicting cooperation relationships in a movie network, and predicting POI (point of interest) visits in a LBSN (location-based social network.) In our first example, a movie network is constituted by multiple types of entities (e.g., movies, participants, studios, and genres) and multiple types of links among these entities. Note that other applications may have similar settings. To clearly represent the semantic meanings in such a network, we utilize the meta-path-based prediction model whose advantages are two-fold. First, the meta-path-based method systematically retrieves topological features in a heterogeneous information network. Second, we use data classification techniques to learn the best weights connected with different topological features in building potential relationships. Empirical studies based on two real datasets from IMDb and Yelp show that the proposed approach is of good prediction quality in practical applications.