M ODELING L ABEL S PACE I NTERACTIONS IN M ULTI LABEL C LASSIFICATION USING B OX E MBEDDINGS

Multi-label classification is a challenging structured prediction task in which a set of output class labels are predicted for each input. Real-world datasets often have taxonomic relationships between labels which can be explicit, implicit, or partially observed. Most existing multi-label classification methods either ignore the label taxonomy or require the complete specification of the taxonomy at training and inference time to enforce coherence in their predictions. In this work we intro-duce the multi-label box model (MBM), a multi-label classification method that combines the encoding power of neural networks with the inductive bias of probabilistic box embeddings (Vilnis et al., 2018), which can be understood as trainable Venn-diagrams based on hyper-rectangles. By representing labels as boxes, MBM is able to capture taxonomic relations among labels without them being provided explicitly. Furthermore, since MBM learns the label-label relationships from data and represents them as calibrated conditional probabilities, it provides a high degree of interpretability. This interpretability also facilitates the injection of partial information about label-label relationships into model training, to further improve its consistency. We provide theoretical grounding for our method and show ex-perimentally the model’s ability to learn the true latent taxonomic structure from data. Through extensive empirical evaluations on twelve multi-label classification datasets, we show that MBM can significantly improve taxonomic consistency while maintaining the state-of-the-art predictive performance. 1