Multi-label Classification of Gene Function using MLPs

This paper describes how a single multi-output MLP can be applied to multi-label classification tasks, and reports on the application of the technique to predicting gene function for arabidopsis - a small flowering plant, and one of the most completely sequenced eukaryotic genomes. Comparison of the classification characteristics of the multi-output MLP with that of multiple binary classifiers reveals several differences, most notably a more rapid fall-off in sensitivity as the output cutoff value is increased. These differences are due to an increased peakedness in the distribution of output values as compared with the distribution of outputs from binary networks. Various explanations are offered to account for this.