Sampling high-quality clicks from noisy click data

Click data captures many users’ document preferences for a query and has been shown to help significantly improve search engine ranking. However, most click data is noisy and of low frequency, with queries associated to documents via only one or a few clicks. This severely limits the usefulness of click data as a ranking signal. Given potentially noisy clicks comprising results with at most one click for a query, how do we extract high-quality clicks that may be useful for ranking? In this poster, we introduce a technique based on query entropy for noise reduction in click data. We study the effect of query entropy and as well as features such as user engagement and the match between the query and the document. Based on query entropy plus other features, we can sample noisy data to 15% of its overall size with 43% query recall and an average increase of 20% in precision for recalled queries. Categories and Subject Descriptors H.3.3 [Information Storage and Retrieval]: Information Se...

• Type: Conference
• Year: 2010
• Where: WWW
• Authors: Adish Singla, Ryen W. White
• Keywords: Click Data, Internet Technology, Query Entropy, Search Engine Ranking, WWW 2010