Sparse Spatial Selection for Novelty-Based Search Result Diversification

Novelty-based diversification approaches aim to produce a diverse ranking by directly comparing the retrieved documents. However, since such approaches are typically greedy, they require O(n) documentdocument comparisons in order to diversify a ranking of n documents. In this work, we propose to model novelty-based diversification as a similarity search in a sparse metric space. In particular, we exploit the triangle inequality property of metric spaces in order to drastically reduce the number of required document-document comparisons. Thorough experiments using three TREC test collections show that our approach is at least as effective as existing novelty-based diversification approaches, while improving their efficiency by an order of magnitude.