Pivot-based approximate k-NN similarity joins for big high-dimensional data

Given an appropriate similarity model, the k-nearest neighbor similarity join represents a useful yet costly operator for data mining, data analysis and data exploration applications. The time to evaluate the operator depends on the size of datasets, data distribution and the dimensionality of data representations. For vast volumes of high-dimensional data, only distributed and approximate approaches make the joins practically feasible. In this paper, we investigate and evaluate the performance of multiple MapReduce-based approximate k-NN similarity join approaches on two leading Big Data systems Apache Hadoop and Spark. Focusing on the metric space approach relying on reference dataset objects (pivots), this paper investigates distributed similarity join techniques with and without approximation guarantees and also proposes high-dimensional extensions to previously proposed algorithms. The paper describes the design guidelines, algorithmic details, and key theoretical underpinnings of the compared approaches and also presents the empirical performance evaluation, approximation precision, and scalability properties of the implemented algorithms. Moreover, the Spark source code of all these algorithms has been made publicly available. Key findings of the experimental analysis are that randomly initialized pivot-based methods perform well with big high-dimensional data and that, in general, the selection of the best algorithm depends on the desired levels of approximation guarantee, precision and execution time.