A Multi-objective Evolutionary Algorithm with Symmetrical Uncertainty Crossover Operator for Feature Selection in Disease Classification

Feature selection plays an important role in the classification and prediction of disease datasets, and it can be established as a multi-objective optimization problem. Nowadays, evolutionary algorithms are a popular toolbox for solving multi-objective feature selection problems. However, for feature selection problems, the Pareto front is sparse. The traditional evolutionary algorithms are prone to getting stuck in local optima, and it is difficult to obtain a set of non-dominated solutions considering both convergence and diversity. In order to solve the above problems, two improvement strategies based on MOEA/D are proposed in this paper. Firstly, to enhance the diversity of the initial population distribution and accelerate the convergence speed of the algorithm, a probability interval-based initialization method is proposed. Secondly, in order to avoid the high similarity between solutions which leads to local optima, a symmetric uncertainty based crossover operator is proposed. According to the symmetric uncertainty between features and class categories, the crossover operator preferentially generates offspring containing the features highly related to classification, which will enhance the classification accuracy of solutions. Finally, three state of the art algorithms are compared by using the disease related datasets, and the results indicate that the proposed algorithm can effectively improve the classification performance.