A parallel processing approach to incremental conceptual clustering

Conceptual clustering summarizes and organizes data allowing for better inference ability. While preferred for their limited memory flavor and adaptability, incremental algorithms are usually bound to sequential execution and little work has been done to explore this artificial limitation. The authors describe a way to map incremental tree-building algorithms on multiprocessors and present an actual implementation, the COBWEB (D.H. Fisher, Machine Learning, vol.2, p.139-72, 1987) based PICC (Parallel Incremental Conceptual Clustering) and its performance. PICC is implemented using the ACTORS-like programming language OODMC on an Encore computer. The MIMD model of computation is analyzed both theoretically and through various experiments Efficiency of multiprocessor use is discussed. The method proves flexible and applicable to a large class of algorithms