Synthetic Sample Generation for Label Distribution Learning

Label Distribution Learning (LDL) is a general learning framework that assigns an instance to a distribution over a set of labels rather than a single label or multiple labels. Current LDL methods have proven their effectiveness in many machine learning applications. As of the first formulation of the LDL problem, numerous studies have been carried out that apply the LDL methodology to various real-life problem solving. Others have focused more specifically on the proposal of new algorithms. The purpose of this article is to start addressing the LDL problem as of the data pre-processing stage. The baseline hypothesis is that, due to the high dimensionality of existing LDL data sets, it is very likely that this data will be incomplete and/or that poor data quality will lead to poor performance once applied to the learning algorithms. In this paper, we propose an oversampling method, which creates a superset of the original dataset by creating new instances from existing ones. Then, we apply already existing algorithms to the pre-processed training set in order to validate the effcacy of our method. The effectiveness of the proposed SSG-LDL is verified on several LDL datasets, showing significant improvements to the state-of-the-art LDL methods.