DOI: 10.5176/978-981-08-7656-2ATAI2010-57
Authors: Xinyi Shu, Yujiu Yang and Wenhuang Liu
Abstract:
Web search click logs, reflecting whether users are satisfied with the search results, are the most extensive and invaluable information resources of user preference. A central problem in click log analysis is to estimate the user perceived relevance of each query-URL pair. Many click model shave been proposed to solve this problem, but they all have a common problem: the examine-next probability only depends on the current result or the preceding last clicked result. Intuitively, whether a user will continue to see the next result is supposed to be determined by whether one is satisfied with the information got from one’s historical clicked results, not only the current result.Therefore, we propose the multi-click dependent model (MCDM) that takes all the preceding clicked results into consideration. In the new model, the examine-next probability is decided by the click variables of each clicked result. We evaluate the proposed model on a real-world data set consisting of about 3.02 million query sessions obtained from a Chinese commercial search engine Sougou to test the performance of MCDM. The experiment results show that MCDM out performs the existing click models in metrics such as log-likelihood, click perplexity, last click prediction error, especially on less-frequent queries and bottom positions of query sessions.
Keywords: Behavior Analysis, Click model, Query logs, Ranking
Updating...