Exploring Feature-Level Duplications on Imbalanced Data Using Stochastic Diffusion Search

Swarm intelligence mimics the behaviours of social insects like bees, wasps and ants to offer powerful problem solving metaheuristic which lies in a network of interactions amongst the agents of a multiagent system as well as with their environment. One of the computer algorithms inspired by swarm intelligence is the stochastic diffusion search (SDS). SDS uses some of the processes and techniques found in swarm to solve search and optimisation problems. In this paper, a hybrid approach is proposed to deal with real-world imbalanced data. The proposed model involves oversampling the minority class, undersampling the majority class as well as optimising the parameters of the classifier, Support Vector Machine (SVM). The proposed model uses Synthetic Minority Over-sampling Technique (SMOTE) to perform the oversampling and the agents of a swarm intelligence technique, SDS, to perform an ‘informed’ undersampling on the majority classes. The use of this swarm intelligence technique in conducting the undersampling tasks is investigated and its impact on improving the classification results is demonstrated. In addition to comparing the agents-led undersampling with random undersampling, the results are contrasted against other best known techniques on nine real-world datasets. Additionally, further experiments are designed to explore the behaviour of the SDS agents during the undersampling process.