Improved Multilabel Classification with Neural Networks

This paper considers the multilabel classification problem, which is a generalization of traditional two-class or multi-class classification problem. In multilabel classification a set of labels (categories) is given and each training instance is associated with a subset of this label-set. The task is to output the appropriate subset of labels (generally of unknown size) for a given, unknown testing instance. Some improvements to the existing neural network multilabel classification algorithm, named BP-MLL, are proposed here. The modifications concern the form of the global error function used in BP-MLL. The modified classification system is tested in the domain of functional genomics, on the yeast genome data set. Experimental results show that proposed modifications visibly improve the performance of the neural network based multilabel classifier. The results are statistically significant.