Multi-Label Classification Based on the Improved Probabilistic Neural Networ

Abstract

This paper aims to overcome the defects of the existing multi-label classification methods, such as the insufficient use of label correlation and class information. For this purpose, an improved probabilistic neural network for multi-label classification (ML-IPNN) was developed through the following steps. Firstly, the traditional PNN was structurally improved to fit in with multi-label data. Then secondly, a weight matrix was introduced to represent the label correlation and synthetize the information between classes, and the ML-IPNN was trained with the backpropagation mechanism. Finally, the classification results of the ML-IPNN on three common datasets were compared with those of the seven most popular multi-label classification algorithms. The results show that the ML-IPNN outperformed all contrastive algorithms. The research findings brought new light on multi-label classification and the application of artificial neural networks (ANNs).