A Bayesian nonparametric model for multi-label learning

Publication Type:
Journal Article
Citation:
Machine Learning, 2017, 106 (11), pp. 1787 - 1815
Issue Date:
2017-11-01
Full metadata record
Files in This Item:
Filename Description Size
ML-final.pdfAccepted Manuscript641.98 kB
Adobe PDF
© 2017, The Author(s). Multi-label learning has become a significant learning paradigm in the past few years due to its broad application scenarios and the ever-increasing number of techniques developed by researchers in this area. Among existing state-of-the-art works, generative statistical models are characterized by their good generalization ability and robustness on large number of labels through learning a low-dimensional label embedding. However, one issue of this branch of models is that the number of dimensions needs to be fixed in advance, which is difficult and inappropriate in many real-world settings. In this paper, we propose a Bayesian nonparametric model to resolve this issue. More specifically, we extend a Gamma-negative binomial process to three levels in order to capture the label-instance-feature structure. Furthermore, a mixing strategy for Gamma processes is designed to account for the multiple labels of an instance. The mixed process also leads to a difficulty in model inference, so an efficient Gibbs sampling inference algorithm is then developed to resolve this difficulty. Experiments on several real-world datasets show the performance of the proposed model on multi-label learning tasks, comparing with three state-of-the-art models from the literature.
Please use this identifier to cite or link to this item: