Interpretable self-labeling semi-supervised classifier
Host Publication: Proceedings of the 2nd Workshop on Explainable Artificial Intelligence
Authors: I. Del Carmen Grau Garcia, D. Sengupta, M. M. Garcia Lorenzo and A. Nowé
Publication Date: Jul. 2018
Semi-supervised classification refers to a type of pattern classification problem involving both labeled and unlabeled data, where the number of labeled instances is often significantly smaller compared to the number of unlabeled ones. Although there exist several semi-supervised classifiers with high performance over different tasks, most of them are complex models that do not allow explaining the obtained outcome, thus behaving like black boxes. In this paper, we perform a critical analysis of the interpretability of state-of-the-art semisupervised classification approaches. In addition, we present a self-labeling grey-box classifier that uses a black-box to estimate the missing class labels and an interpretable white-box to make the actual predictions. The main contribution of this model relies on its transparency while also being able to outperform most state-of-the-art semisupervised classifiers.