Weighted Information Gain and User Clicks on Web Search Results Yun Zhou , W.Bruce Croft, Department of Computer Science, University of Massachusetts Amherst, 140 Governor Dr. , Amherst.MA,01002, USA {yzhou, croft}@cs.umass.edu Abstract. In this poster, we demonstrate that the WIG (Weighted Information Gain) technique, originally proposed for retrieval performance prediction and shown to be effective particularly in Web search environments, has an interesting connection with user clicks on Web search results. Specifically, we observe that high WIG scores generally suggest more clicks. This makes WIG a useful feature for predicting user’s preference for search results, which has potential applications in many important areas such as the automatic tuning of search engine parameters, personalization, sponsored search and others. Keywords: WIG, click, prediction, user preference 1 Introduction WIG (Weighted Information Gain) was demonstrated as an effective technique for retrieval performance prediction in Web search environments [1]. A significant correlation between the WIG score and retrieval performance was observed in various Web search scenarios. However, the evaluation of WIG in [1] was performed under laboratory settings which consist of carefully-selected topic sets with relevance judgments made by human assessors and pre-defined retrieval tasks. The high cost of producing relevance judgments makes it difficult to test the performance of WIG in real-world Web search. Instead of relying on human relevance judgments, we make use of user clicks on search results to approximate relevance judgments. In fact, the click of a document can be viewed as an implicit and rough relevance feedback made by the user. Although it can be noisy, click information, when used in aggregation, is a good indicator of relevance [2]. In this poster, we experiment on realistic Web data with click information gathered from a commercial search engine. Consistent with the finding in [1] that high WIG scores generally correspond to high retrieval performance, we observe a tendency that high WIG scores predict more clicks on search results. This confirms earlier experiments showing that WIG is a useful feature for predicting the outcome of a query, which has potential applications in many important areas such as the automatic tuning of search engine parameters, personalization, sponsored search and others.