Tuning Schema Matching Systems Using Genetic Algorithms

Most recent schema matching systems combine multiple components, each of which employs a particular matching technique with several knobs. The multi-component nature makes matching systems extensible and customizable, however, it also brings tuning prob-lem that is to determine which components to execute and how to adjust the knobs of these components, for given matching situation.In this thesis, we start by surveying the existing solutions to the tuning problem. Based on the insights about the state of the art, we have developed GATuner which generates train-ing data set and tunes schema matching systems automatically. We promote the performance of GATuner by implementing fine-grained parallel genetic algorithms on CUDA. Our ap-proach has the following interesting features.1. Provide a matching scenario generator which automatically produces matching sce-narios for input schemas. It handles both relational and XML schemas, and it can be ex-tended to other data representations by developing certain perturbations.2. Tune schema matching systems using genetic algorithms (GAs). GAs generate so-lutions to optimization problems using techniques inspired by natural evolution. Compared with other approaches, GATuner is more likely to find the global optimum of tuning prob-lem.3. Promote the performance of GATuner by implementing fine-grained parallel genetic algorithms (PGAs) on CUDA. Our approach could effectively improve the performance of GATuner with very low extra cost.Experiments demonstrate that GATuner provides more qualified matches over different domains, and the parallelization effectively improves the performance especially when large population involved.