PhD Proposal: Collective Multi-Relational Network Mining

Our world is becoming increasingly interconnected, and the study of networks and graphs are becoming more important than ever.
While the study of networks is not new, until recently most traditional research in this domain have focused on networks with a single node type and a single edge type. However, most networks are formed between different types of nodes and contain different types of links. Predictive models can improve their performances by leveraging the multi-relational and heterogeneous network structures. Another important factor in the network domain is entity interactions. Networks are natural domains for collective and joint predictions, where inferred information about one entity should change the model's belief about other related entities. In this thesis, I propose models that can effectively leverage the multi-relational nature of networks and collectively make predictions on links and nodes.
In the first part of this proposal, I present models for making predictions on links in an augmented bipartite multi-relational structure for drug-target interaction networks, and later extend them to recommender systems domain. In the second part, I present an approach to make predictions on nodes for spammer detection in evolving multi-relational social networks based on the network structural features and the sequences of edge generation. I then present a model to collectively determine the credibility of users as well as their probabilities of being spammers. For my future work, I plan to extend my studies to scale the proposed link prediction model via effective similarity selection. I then plan to study methods to incorporate continuous contextual information to make better predictive models for link prediction and recommender systems.
Examining Committee:
Committee Chair: - Dr. Lise Getoor
Dept's Rep. - Dr. Larry Davis
Committee Member(s): - Dr. Hal Daume III