Similarity Joins of Sparse Features

Identifying all pairs of records from two datasets whose similarity exceeds a given threshold is crucial for data cleaning and clustering. Our work on similarity-joins is motivated by detecting fraud and abuse. We focus on similarity-joins of sparse features, where records represent sparse sets, multisets, or vectors. Most state-of-the-art techniques are distributed versions of sequential algorithms. This is the reason they fail to scale when the alphabet is large, the records are large, or some elements are shared by numerous records. In this paper, we propose FastScalableSparseJoiner (FSSJ) that introduces quasi-prefix filtering, a novel flavor of prefix filtering [7] that exploits the skew in element popularity to avoid processing the most-popular elements without broadcasting the sorted elements to all executors. FastScalableSparseJoiner effectively prunes candidate pairs, and efficiently distributes the computation of partial results for load-balancing across executors. FSSJ can be adopted in any shared-nothing architecture. Based on our evaluation on synthetic and real datasets using a Spark-based implementation, FSSJ is very competitive on small datasets, and is the only algorithm that can join industry-scale datasets even with limited resources.