Laplacian-based Semi-supervised Multi-Label Regression
Vivien Kraus, Khalid Benabdeslem, Bruno Canitia · 2020
In multi-label learning, one has to find a model suitable for predicting multiple values for the same individual, based on the same features. The effectiveness of most multi-label algorithms lies in the fact that it is able to consider the correlations between the related labels. On the other hand, in many applications, a multi-label learning task incurs a high cost for the annotation of a single data point. This leads to a dataset consisting of a few labeled data points, and many more unlabeled data points. In this scenario, semi-supervised methods can take advantage of the unlabeled data points. In this article, we propose a new algorithm for multi-label semi-supervised regression, LSMR, as a multi-label extension of a semi-supervised regression algorithm. We provide experimental results on some publicly-available regression datasets showing the effectiveness of our approach.