A PSO algorithm for multi-objective cost-sensitive attribute reduction on numeric data with error ranges

Multi-objective cost-sensitive attribute reduction is an attractive problem in supervised machine learning. Most research has focused on single-objective minimal test cost reduction or dealt with symbolic data. In this paper, we propose a particle swarm optimization algorithm for the attribute reduction problem on numeric data with multiple costs and error ranges and use three metrics with which to evaluate the performance of the algorithm. The proposed algorithm benefits from a fitness function based on the positive region, the selected n types of the test cost, a set of constant weight values \(w_{i}^k\), and a designated non-positive exponent \(\lambda \). We design a learning strategy by setting dominance principles, which ensures the preservation of Pareto-optimal solutions and the rejection of redundant solutions. With different parameter settings, our PSO algorithm searches for a sub-optimal reduct set. Finally, we test our algorithm on seven UCI (University of California, Irvine) datasets. Comparisons with alternative approaches including the \(\lambda \)-weighted method and exhaustive calculation method of reduction are analyzed. Experimental results indicate that our heuristic algorithm outperforms existing algorithms.