An Approach to Frequent Pattern Discovery from Gene Expression Data Using PSO Variants

Abstract

Pattern mining has always attracted a huge attention for generation of large amount of patterns and association between them. Though it's one of the major data mining tasks but it has always been a time consuming process as a large scale of patterns and associations rules gets generated. To reduce the time of consumption it was preferable to discretize the data matrix in the range of 0 to 1 and for this the fuzzy the membership function has been used which is quite simple in its concept and strategy. Owing to the concept of fuzzy logic, certain evolutionary algorithms (EAs) also gained popularity to optimize the process of mining patterns from the fuzzy sets. For this, Particle swarm optimization (PSO) was used which is supposed to provide better results as compared to other EA like genetic algorithm, ant colony optimization etc. But it was found that there are certain versions of PSO that provided much better results than the standard PSO algorithm. In this paper, the gene expression data set was fuzzified for the purpose of discretization in the range of 0 to 1. A Frequent Pattern (FP) growth algorithm was used to generate set of frequent patterns. These patterns were used as the initial population and the mean squared residue (MSR) score was used as an evaluation criteria. Fully Informed Particle Swarm Optimization (FIPSO), Dynamic Multi Swarm Particle Swarm Optimization (DMS-PSO), Comprehensive Learning Particle Swarm Optimization (CLPSO), Vector Evaluated Particle Swarm Optimization (VEPSO) etc are the certain versions of PSO that were used and they provided much better results as compared to standard PSO algorithm. But the VEPSO algorithm outperformed the other three algorithms in terms of generation of best individual frequent patterns, runtime and the volume of mean squared residue (lower the MSR score the better is the quality of the patterns).