STRUCTURAL TOPIC MINING IN WEB COLLECTIONS

This paper introduces structural topic mining: an approach for discovering and describing thematically related document groups in large document collections. A collection is viewed as a directed graph where vertices represent documents and arcs represent connections among these. Because a document is likely to have more connections to documents of the same theme, we have assumed that topics have the structure of a graph cluster, i.e. a group of vertices with more arcs to the inside of the group and fewer arcs to the outside. So, topics could be discovered by clustering the document graph; a local approach is used for scalability. We also extract properties (keywords and representative documents) from clusters. This approach was tested over Wikipedia, and the resulting clusters in fact correspond to topics; this shows that topic mining can be treated as a graph clustering problem. Comparative results suggest considerable quality at a low cost.