Parallel out-of-core divide-and-conquer techniques with application to classification trees

Classification is an important problem in the field of data mining. Construction of good classifiers is computationally intensive and offers plenty of scope for parallelization. Divide-and-conquer paradigm can be used to efficiently construct decision tree classifiers. We discuss in detail various techniques for parallel divide-and-conquer and extend these techniques to handle efficiently disk-resident data. Furthermore, a generic technique for parallel out-of-core divide-and-conquer problems is suggested. We present pCLOUDS, the parallel version of the decision tree classifier algorithm CLOUDS, capable of handling large out-of-core data sets. pCLOUDS exhibits excellent speedup, sizeup and scaleup properties which make it a competitive tool for data mining applications. We evaluate the performance of pCLOUDS for a range of synthetic data sets on the IBM-SP2