Vpso-based ccr-elm for imbalanced classification

Abstract

In class-specific cost regulation extreme learning machine (CCR-ELM) for the class imbalance problems, the key parameters, including the number of hidden nodes, the input weights, the hidden biases and the tradeoff factors are normally chosen randomly or preset by human. This made the algorithm responding slowly and generalization worse. Unsuitable quantity of hidden nodes might form some useless neuron nodes and make the network complex. So an improved CCR-ELM based on particle swarm optimization with variable length is present. Each particle consists of above key parameters and its length varies with the number of hidden nodes. The experimental results for nine imbalance dataset show that particle swarm optimization with variable length can find better parameters of CCR-ELM and corresponding CCR-ELM had better classification accuracy. In addition, the classification performance of the proposed classification algorithm is relatively stable under different imbalance ratios.