A Comparison Study of Strategies for Combining Classifiers from Distributed Data Sources

Distributed data mining (DDM) is an important research area. The task of distributed data mining is to extract and integrate knowledge from different sources. Solving such tasks requires a special approach and tools, different from those applied to learning from data located in a single database. One of the approaches suitable for the DDM is to select relevant local patterns from the distributed databases. Such patterns often called prototypes, are subsequently merged to create a compact representation of the distributed data repositories. Next, the global classifier, called combiner, can be learned from such a compact representation. The paper proposes and reviews several strategies for constructing combiner classifiers to be used in solving the DDM tasks. Suggested strategies are evaluated experimentally. The evaluation process is based on several well-known benchmark data sets.