User profiling in social networks based on graph data: bridging local and global structures

User profiling plays a key role in adaptive systems on online social networks (OSN). Building user profiles allows to identify user's interests (interest modeling) and predict their future behaviors (behavioral modeling). While various data sources can be exploited on OSNs for user profiling, graph data is a more structured and more abundant data source compared to other data sources like content data. In fact, graphs can be used to capture different types of interactions and relationships among users. By exploiting graph data, users can be profiled based on their neighbors. In this thesis, we aim to address different profiling problems based on graph data, including interest modeling and behavioral modeling, where the main focus is on integrating local and global structures. In fact, to successfully exploit graph data for user profiling, both local and global structures need to be analyzed, i.e., profiling a node only based on its immediate neighbors (local structures) is not fully effective. Such models miss the precious information that can be exploited by analyzing the entire network (global structures). However, as the network can be large, and the connection between nodes may be complex, building efficient and effective models that can bridge local and global structures is challenging. We define novel interest profiling problems on interest-based networks that require us to bridge local and global structures efficiently. In particular, we focus on followee/follower networks in microblogging websites as interest-based networks. To address the efficiency challenge of interest profiling, we propose the notion of profiling based on hub nodes. The hub nodes can be viewed as the cluster centers of the network where each cluster depicts a certain interest topic. Our core idea is that users can efficiently be profiled by transitively profiling them based on the hub nodes they are connected to. We also investigate user behavior modeling on networks that represent user's behavioral interactions. Various methods have been introduced for user modeling on classic networks with a single type of interaction that is able to bridge local and global structures. However, their effectiveness is limited on networks with multi-types of interactions due to the complexity of relations on such networks. We introduce models to build behavioral models in networks with multi-types of interactions that systematically address the complexity challenge