An Execution Cost Model for Aggregate Queries over Knowledge Graphs

Abstract

Recent influx of knowledge graphs (KGs) focuses on managing large-scale and real-world facts as a big graph. Querying KGs is essential for a wide range of applications and has been made remarkable progress. One of the most most important query forms for KGs is aggregate query (e.g., "what is the average price of cars produced in Germany?"). Users can leverage aggregate query on KGs to obtain potential statistical information of interest. The latest efficient and effective solution for answering aggregate queries on KGs is Approximate aggregate Queries with semantic-aware Sampling (AQS), which can achieve a trade-off between query time and result accuracy. The basic idea is concluded as: estimate an approximate aggregate result based on a random sample collected from a KG and guarantee that the relative error of approximate result is bounded by a pre-defined error bound. This solution performs well for simple aggregate queries but suffers from the complex aggregate queries having multiple sub-queries, such as star queries. This is because it’s non-trivial to configure an appropriate pre-defined error bound for each sub-query. If we set a large error bound for each sub-query, then we can quickly obtain the complex query’s result but having low quality. Otherwise, we can obtain the high quality of the complex query’s result but consuming long time for small pre-defined error bound. Obviously, it has a tradeoff between the error bound and query time. In this paper, we propose an execution cost model for AQS, which can accurately describe the functional relationship between query time and the pre-defined error bound, and help to determine an appropriate pre-defined error bound for each sub-query of a complex query. Extensive experiments over real KGs confirm the effectiveness and efficiency of our cost model.