Feature relation network that can identify underlying data structure for effective pattern classification

This paper proposes a feature relation network (FRN) to model the underlying feature relation structures of a set of observations. A pattern classification system is then constructed based on the feature relation network, namely PCS-FRN. During training process, PCS-FRN will form an attractor for each group of samples in order to lower the overall energy states. The attractor, or a feature relation network, reflects the underlying data structure that can discriminate different classes. Parameters of PCS-FRN are estimated by the multi-dimensional evolutionary algorithm. The PCS-FRN system was tested on a synthetic dataset and three real-world medical datasets and compared with conventional classification techniques. Experiment results show that PCS-FRN can achieve better classification accuracies on both binary and multi-class problems.