A new parallelization of subgraph isomorphism refinement for classification and retrieval of conceptual structures

Abstract

Major applications of conceptual structures will require quick response times on extremely large knowledge bases. Although algorithmic developments have provided tremendous improvements in speed, we believe implementation on parallel processors will be needed to meet long-term needs. This paper presents a new parallelization of a subgraph isomorphism refinement algorithm for performing projection tests and retrieving conceptual structures (CS). The improved algorithm is faster, requires fewer processors, and is compatible with recent relation-based representations of CS. The new parallelization takes advantage of the features of contemporary massively parallel machines by exploiting bit-parallelism in the data words. Processing numerous CS on a single parallel array using load balancing integrated with multi-level indexed search, it combines the strengths of prior parallel subgraph isomorphism parallelizations. It incorporates lattice codes of the concept-type hierarchy, forming all node candidate binding lists in parallel. Simulation results of the behavior of the refinement algorithm with parameterized synthetic data sets are presented.