Improving website search with server log analysis and multiple evidence combination

Abstract

Despite the success of global search engines, website search is still problematic in its retrieval accuracy. Server logs contain a rich source of information about how users actually access a website. In this paper, we propose a novel approach of using server log analysis to extract terms to build the web page representation, which is a new source of evidence for website search. Then, we use multiple evidence combination to combine this log-based evidence with text-based and anchor-based evidence. We test the performance of different combination approaches the combination of representations and of ranking scores, using linear combination and inference network models. We also consider different baseline retrieval models. Our experimental results have shown that the server log, when used in multiple evidence combination, can improve the effectiveness of website search, whereas the impact on different models is different.