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.

PDF file

In: World Wide Web Conference

Details