Fusion and diversification in information retrieval

Data fusion and search result diversification are two critical research topics in information retrieval. Data fusion approaches combine search result lists in order to produce a new and hopefully better ranking. We propose two data fusion models for microblog search that exploit temporal information and infer rank scores of missing documents in the lists to be fused. We also propose a fusion method based on manifolds. The method constructs manifolds, let low ranked documents be rewarded to be relevant by high ranked documents in the same manifolds, and utilize the top-k documents as anchors to enhance the efficiency of data fusion. Search result diversification is widely being studied as a way of tackling query ambiguity. Instead of trying to identify the "correct" interpretation behind a query, the idea is to make the search results diversified so that users with different backgrounds will find at least one of these results to be relevant. We examine the hypothesis that data fusion can improve performance in terms of diversity metrics, and proposes a new data fusion method, called diversified data fusion for search result diversification. We also study the problem of personalized diversification via supervised learning, with the goal of enhancing both diversification and personalization performance. The results in this thesis show how both our proposed data fusion and search result diversification methods improve retrieval performance and how they relate to each other. The insights in this thesis may be used to improve retrieval performance for a range of tasks in information retrieval.