Abstract

A bipartite query-URL graph, where an edge indicates that a document was clicked for a query, is a useful construct for finding groups of related queries and URLs. Here we use this behavior graph for classification. We choose a click graph sampled from two weeks of image search activity, and the task of “adult” filtering: identifying content in the graph that is inappropriate for minors. We show how to perform classification using random walks on this graph, and two methods for estimating classifier parameters.