cluTM: Content and Link Integrated Topic Model on Heterogeneous Information Networks

Abstract

Topic model is extensively studied to automatically discover the main themes that pervade a large and unstructured collection of documents. Traditional topic models assume the documents are independent and there are no correlations among them. However, in many real scenarios, a document may be interconnected with other documents and objects, and thus form a text related heterogeneous network, such as the DBLP bibliographic network. It is challenging for traditional topic models to capture the link information associated to diverse types of objects in such a network. To this end, we propose a unified Topic Model cluTM by incorporating both the document content and various links in the text related heterogeneous network. cluTM combines the textual documents and the link structures by the proposed joint matrix factorization on both the text matrix and link matrices. Joint matrix factorization can derive a common latent semantic space shared by multi-typed objects. With the multi-typed objects represented by the common latent features, the semantic information can be therefore largely enhanced simultaneously. Experimental results on DBLP datasets demonstrate the effectiveness of cluTM in both topic mining and multiple objects clustering in text related heterogeneous networks by comparing against state-of-the-art baselines.