A Neural Networks Approach to SPARQL Query Performance Prediction

Abstract

The SPARQL query language is the standard for querying RDF data and has been implemented in a wide variety of engines. These engines support hundreds of public endpoints on the Web which receive thousands of queries daily. In many cases these endpoints struggle when evaluating complex queries or when they receive too many of them concurrently. They struggle mostly since some of these queries need large amounts of resources to be processed. All these engines have an internal query optimizer that proposes a supposedly optimal query execution plan, however this is a hard task since there may be thousands of possible query plans to consider and the optimizer may not chose the best one. Herein we propose the use of machine learning techniques to help in finding the best query plan for a given query fast, and thus improve the SPARQL servers' performance. We base such optimization in modeling SPARQL queries based on their complexity, operators used within the queries and data accessed, among others. In this work we propose the use of Dense Neural Networks to improve such SPARQL query processing times. Herein we present the general architecture of a neural network for optimizing SPARQL queries and the results over a synthetic benchmark and real world queries. We show that the use of Dense Neural Networks improve the performance of the Nu-SVR approach in about 50% in performance. We also contribute to the community with a dataset of 19,000 queries.