Fast Similarity Search with the Earth Mover's Distance via Feasible Initialization and Pruning

摘要

The Earth Mover's Distance (EMD) is a similarity measure successfully applied to multidimensional distributions in numerous domains. Although the EMD yields very effective results, its high computational time complexity still remains a real bottleneck. Existing approaches used within a filter-and-refine framework aim at reducing the number of exact distance computations to alleviate query time cost. However, the refinement phase in which the exact EMD is computed dominates the overall query processing time. To this end, we propose to speed up the refinement phase by applying a novel feasible initialization technique (INIT) for the EMD computation which reutilizes the state-of-the-art lower bound IM-Sig. Our experimental evaluation over three real-world datasets points out the efficiency of our approach (This work is partially based on [12]).