Investigating Data Distribution for Classification Using PSO with Adversarial Network

Xiaojing Zhang, Jiawei Fan, Lin Wang, Bo Yang

Abstract: Data distribution plays a key role in the performance of various classification algorithms. Artificial neural network (ANN) has been widely applied in considerable complex tasks because of its excellent universal approximation capability. Although the floating centroids method (FCM) provides an effective and diverse output encoding that removes the fixed centroids constraint, the adaptive mechanism between the FCM-based neural network classifier and the data distribution has not been studied. In this paper, we design an adversarial network to investigate the characteristics of FCM and adopt a particle swarm optimization to evolve the data distribution. Experimental results demonstrated that FCM show the characteristics of diversified centroids, flexible clustering, and insensitivity to data scale, thereby obtaining competitive performance.