Click-graph modeling for facet attribute estimation of web search queries

We use clickthrough data of a Japanese commercial search engine to evaluate the similarity between a query and a facet category from the patterns of clicks on URLs. Using a small number of seed queries, we extract a set of topical words forming search queries together with the same facet directive words, e.g., 'recipe' in 'curry recipe' or 'apple pie recipe'. We used a PageRank-like random walk approach on query-URL bipartite graphs called "Biased ClickRank" to propagate facet attributes through click bipartite graphs. We noticed that queries to URL links are too sparse to capture query variations whereas queries to domain links are too coarse to discriminate among the different usages of broadly related queries. We introduced edges and vertices corresponding to the decomposed URL paths into the click graph to capture the click pattern differences at an appropriate granularity level. Our expanded graph model improved recalls as well as average precision against baseline graph models.