Predicting Users’ Search Behavior Using Stochastic Multi-mode Network Models

Abstract

Multidimensional relationships can be represented as a multi-mode network or graph, where each vertex or node corresponds to an object, and each edge or link is attributed to one of the multiple types of relationships between a pair of objects. Web search log includes users' search behavior and can also be represented as such a multi-mode network, where each vertex corresponds to a query and each attributed edge corresponds to a relationship between a pair of queries. The relational attributes can be derived from multiple assumptions, for instance, two queries are considered to be related to each other when two different users input those queries and click through from respective search result lists to the sameWeb pages. In order to analyze such complex data, this paper proposes a new multi-mode block model based on latent variable modeling. We evaluate the effectiveness of our multi-mode block model through experiments on the task of predicting queries related to each given query using real search query log.