Embedded binary PSO integrating classical methods for multilevel improved feature selection in liver and kidney disease diagnosis

摘要

Feature selection is an important pre-processing technique in the field of data mining. This process removes irrelevant data thereby reduces the number of features. This paper presents a new algorithm called embedded binary particle swarm optimisation (BPSO) to improve the performance of BPSO for feature selection. Embedded BPSO incorporates classical methods to select elite feature subset. The population is refined or extended at regular intervals using the best features from sequential forward selection and sequential backward selection methods. In this study, probabilistic neural network and support vector machine with threefold cross-validation are used to evaluate the particles. The embedded algorithm is verified in the feature selection module of the liver and kidney cancer diagnostic system. The elite features extracted from wrapper based embedded algorithm are used to characterise diseases using the classifier. Findings show that the proposed system is proficient in selecting the best features with minimum error rate.