Index-Based R-S Similarity Joins

Abstract

Similarity Joins are some of the most useful and powerful data processing operations. They retrieve all the pairs of data points between different data sets that are considered similar within a certain threshold. This operation is useful in many situations, such as record linkage, data cleaning, and many other applications. An important method to implement efficient Similarity Joins is the use of indexing structures. The previous work, however, only supports self joins or requires the joint indexing of every pair of relations that participate in a Similarity Join. We present an algorithm that extends a previously proposed index-based algorithm (eD-Index) to support Similarity Joins over two relations. Our approach operates over individual indices. We evaluate the performance of this algorithm, contrast it with an alternative approach, and investigate the configuration of parameters that maximize performance. Our results show that our algorithm significantly outperforms the alternative one in terms of distance computations, and reveal interesting properties when comparing execution time.