Exploit Hierarchical Label Knowledge for Deep Learning

In this paper we propose a methodology based on a complex deep learning network topology, named Hierarchical Deep Neural Network (HDNN), applied to eXtreme Multi-label Text Classification (XMTC) problem. The HDNN topology reproduces the label hierarchy. The main idea arises directly from the assumption that, if the label-set structure is defined, forcing this information into the network topology could improve classification performances and results interpretation. In this way, we define a method to force prior knowledge into the DNN. We perform the experimental assessment on a XMTC task related to a real application domain problem, namely the automatic labelling of biomedical scientific literature extracted from the PubMed. The obtained preliminary results show that, despite the very high computational time needed to update the network weights, a slight performance improvement is obtained, with respect to a classical approach based on Convolution Neural Network (CNN). Some considerations will be drawn out to figure out possible key readings.