A Community-Enhanced Retrieval Model for Text-Rich Heterogeneous Information Networks

In recent years, we have witnessed a large increase in the number of available text-rich heterogeneous networks, in which text documents, users and other objects are interconnected in various ways. Examples include social networks, bibliographic information networks, and collaboration networks. Text-rich heterogeneous networks contain text nodes (e.g., user comments in social networks; scientific publications in bibliographic networks), non-text nodes (e.g., users and companies in social networks), as well as rich relationships between different types of nodes. An important task, which is very useful on its own right, as well as often serves as a preprocessing to mining and analytical tasks, is identifying relevant text information from the heterogeneous information networks given user queries. Unlike Web search and many related problems, the effects of the underlying semantically meaningful patterns in heterogeneous networks play an important role in determining relevant answers for user search queries. For example, there are several structural patterns we can explore to more effectively identify relevant text nodes for a given query. The relevant nodes are likely to come from the same user-topic communities as the query. In addition, the relevant nodes on the same query topic tend to form tight local relationships with each other in the same user-topic community. In this paper, we present a unified and principled framework that effectively integrates network community and local relationship detection with retrieval model construction in a mutually enhancing manner, which leads to a community-enhanced retrieval model for text-rich heterogeneous information networks. Experimental results show that this new model significantly outperforms state-of-the-art baselines.