Improved information retrieval through set-based preference learning

I present a new information retrieval framework based on set-based preference learning that provides users with individually customized search results. While newer information retrieval algorithms can provide a much richer result set than traditional term-weighting methods, they adopt a one-size-fits-all approach. If one user prefers results with a wide variety of content and another prefers results narrowly focused around their query, one or both may be left dissatisfied if they receive the same set of results. I examine the sets of documents that a given user selected in response to previous searches. Using this data, I estimate their preference for result sets in terms of two criteria: the relevance of individual documents to the user's query and the overall diversity of content throughout the result set. Using this learned estimation, I then select a subset of all search results that best balances their personal preference for the two criteria. I demonstrate the effectiveness of my approach on toy domains, data from the Reuters news service, and the 20 Newsgroups data set.;I also present a heuristic to automatically estimate a user's valuation of document relevance relative to document diversity. Many existing methods for diversifying search result sets include parameters that control the role that relevance and diversity play in selecting search results. However, these methods do not automatically customize the values of these controlling parameters to individual users. I present an approach to estimate each user's personal valuation of relevance and diversity in search results automatically.