User Modeling Via Social Network Structure

In recent years, online social networking sites have gained increasing popularity, and become an essential part of people’s life. While users interact with each other on social networks, certain traits are disclosed, e.g., interests, hobbies, mental status, and behavior patterns. Mining and understanding the traits is necessary for improving on-line social networking sites and related services. User modeling, which represents and infers user’s traits, is an essential building block for user understanding. User modeling involves building models to represent user data and designing algorithms to extract and infer the traits.The interactions among users play an important role in online social networks and differs social network with other online services. Interactions among users are the main source of the activities in social network and users interact via their social ties. We represent social ties with a graph, and refer it to social network structure. Previous works have shown that network structure is correlated with user’s certain traits, so it is possible to model users via social network structure. While different online social networking sites focus on various aspects of users, the users are always connected with social ties. Social ties provide the mechanism to share and spread information. It has also been shown that different social networks share certain similar properties. This indicates it is possible to apply a user modeling technique to distinct social networks. In this dissertation, we focus on user modeling using social network structure.User modeling is usually data-driven. The availability of data determines the ca-pacity of user modeling. However, an individual social network can only provides data of a certain aspect of users. In order to enrich the diversity of data, we also studied the cross-domain linking problem. Cross-domain linking aims to link the accounts of a specified user across different sites. We proposed a node similarity and algorithms that utilize the structure and descriptive attributes of social networks for cross-domain linking. We analyzed the property of the proposed similarity and evaluated it in the scenario of de-anonymization. With the proposed method, we won the WSDM 2013 Data Challenge.User modeling is also highly application-oriented. The traits involved in user mod-eling can be categorized as personal traits and social traits. In this dissertation, we studied two instances of the both categories, namely carefulness and social status. We illustrate and discuss the process of user modeling with two instances.We noticed that some users are very careful when establishing new social ties, while the others are somewhat casual. We propose the carefulness and a supervised learning algorithm to capture such behaviors. We evaluated our method with real-world datasets, and found that our model indeed captured the carefulness of users. We then in-corporate carefulness with existing methods for spammer detection and link prediction. Experiments showed that carefulness is able to improve the both applications.We also studied modeling social status with social network data. Status is implic-itly specified in the structure of social networks. We try to infer the status and compare it with social status in the real world. To avoid the noise and bias in social network data, we proposed a membership inference algorithm, which combines ideas of propagation and supervised random walk. Our result shows that it is possible to infer social status from social networks, but certain subtle biases also exist.