A Machine Learning Approach to SPARQL Query Performance Prediction
4 likes1,288 views
This document discusses a machine learning approach to predicting the performance of SPARQL queries over Linked Data without using statistics about the underlying RDF data. It extracts features from SPARQL queries to represent them for machine learning algorithms. The features include algebra features from the SPARQL query expressions and graph pattern features that model the query pattern. Experiments on DBpedia data show the approach can highly accurately predict execution times for common Linked Data queries by training machine learning models on previously executed queries. Future work may incorporate additional features like bandwidth and optimize queries for Linked Data applications and query processing.
A Machine Learning Approach to SPARQL Query Performance Prediction
- 1. A Machine Learning Approach to SPARQL Query Performance Prediction Rakebul Hasan Wimmics Research Team INRIA Sophia Antipolis France
- 2. Context 2 Slide derived from Andreas Blumauer’s Linked Data slides • Linked Data Principles 1. Use URIs as names for things. 2. Use HTTP URIs, so that people can look up those names. 3. When someone looks up a URI, provide useful information, using the standards (RDF, SPARQL). 4. Include links to other URIs, so that they can discover more things.
- 3. Context • W3C Linking Open Data (LOD) Initiative • An initiative to publish open data as Linked Data • From 2 billion triples in 2007 to 30 billion triples in 2011 • Accessing Linked Data – Dereferencing URIs – SPARQL Endpoints 3
- 4. Context • Querying Linked Data – SPARQL Endpoints: SPARQL query service via HTTP implementing SPARQL Protocol – 68% of the data sets provide SPARQL Endpoints as of September 2011 – As of Today, 98% of the triples in LOD cloud are accessible via SPARQL • 57,856,463,005 out of 58,882,358,557 triples http://stats.lod2.eu/ 4
- 5. Context • Understanding Query Behavior in the context of Linked Data – Workload allocation to ensure specific QoS requirements are met – Predicting query performance metrics 5
- 6. Query Performance Prediction • Traditional approaches use underlying data statistics-based cost models to predict query performance • Data statistics are often missing in the Linked Data scenario – Only 32.20 % (95 out of 295) data sources provide a voiD description. • Basic statistics such as number of triples, often not detailed enough for statistics based models – In fact, what makes effective statistics for query cost estimation on RDF is unclear. • Challenge – How to predict query performance without using data statistics? 6
- 7. Understanding performance of database queries • Ganapathi et al. predicting performance metrics of database queries prior to query execution using machine learning. • Akdere et al. use machine learning for predicting query execution time. Ganapathi et al.: Predicting multiple metrics for queries: Better decisions enabled by machine learning, ICDE’09 Akdere et al, Learning-based query performance modeling and prediction, ICDE’12, 7
- 8. Predicting Query Performance • Learn query performance from already executed queries • Challenge: how to model SPARQL query characteristics for machine learning algorithms - feature extraction? 8
- 9. Modeling SPARQL Query Execution • Two types of features – Algebra features: extracted from SPARQL algebraic expression of a query – Graph pattern features: a vector representation of the query pattern of a query relative to the training queries 9
- 10. Modeling SPARQL Query Execution • Algebra features – Jena API to extract SPARQL algebra expressions 10
- 11. • Graph pattern features – Find landmarks in training queries by clustering • K-medoids with approximate graph edit distance – Compute distance between landmark queries and the query in examination to construct a graph pattern feature vector • Approximate graph edit distance for distance computation 11
- 12. Graph Edit Distance • Minimum amount of distortion needed to transform one graph to another – Bipartite matching based approximated graph edit distance with • Previous research shows accurate results with classification problems Riesen et al. “A Novel Software Toolkit for Graph Edit Distance Computation”, 9th IAPR-TC-15, GbRPR 2013 12
- 13. Experiments • 1260 training, 420 validation, and 420 test queries generated from DBPSB benchmark query templates – DBPSB templates cover most commonly used SPARQL query features in the queries sent to DBPedia • DBpedia as RDF data set • Predicting query execution time – k-NN regression with k-D tree – SVM with nu-SVR for regression 13 DBpedia: http://dbpedia.org/ DBPSB: http://aksw.org/Projects/DBPSB.html
- 15. Algebra and graph pattern features 15
- 16. Time Required for Training and Predictions 16
- 17. Summary • Understanding SPARQL query behavior in the Linked Data scenario – Predicting query performance metrics • learn query execution times from already executed queries – without using statistics about the underlying RDF data. – Modeling (vector representation) SPARQL queries for machine learning algorithms • Feature extraction – Highly accurate predictions for common Linked Data queries 17
- 18. Future Work on QPP • Incorporating bandwidth related features. • Query optimization for Linked Data applications: – in place of selectivity estimation for alternative queries? • How to accurately predict performance for single triple patterns – Alternative query construction for Linked Data applications – join order optimization. E.g. Federated Query Processing over Linked Data • How to generate training queries? – Next slide 18
- 19. Statistical Analysis of Query Logs • Approach to systematically generating training queries Mario Arias, Javier D. Fernández, Miguel A. Martínez-Prieto, Pablo de la Fuente: An Empirical Study of Real-World SPARQL Queries, 1st International Workshop on Usage Analysis and the Web of Data, co-located with the 20th International World Wide Web Conference (WWW2011) 19
- 20. • Thank you 20
Editor's Notes
- #3: Web of Documents -> documents were described using HTML and globally identified using URLs Retrieval mechanism: HTTP protocol all these ensured creating a single global data space Data on the Web Many formats and APIs Proprietary interfaces No single global data space – no hyperlinks between data items within different data sources Web of Data -> a single global data space using RDF to publish data on the Web links between data items within different data sources
- #7: Histograms -> on which to create histograms for effective estimation
- #13: The graph edit distance between two graphs is the minimum amount of distortion needed to transform one graph to another. The minimum amount of distortion is the sequence of edit operations with minimum cost. The edit operations are deletions, insertions, and substitutions of nodes and edges.
- #20: Refining training queries from query logs by considering the statistically significant characteristics Bootstrapping: Starting with a initial set of properties, resources and literals and than generate training queries by permutations and combinations of the statistically significant features Simplifying the pattern features: join features, triple pattern features, pattern graph features to represent query patterns