Semi-supervised SPO tree classifier based on the DPC framework

Decision tree is a simple, effective and interpretable algorithm, which has been widely used in different machine learning applications. Recently, the decision tree algorithms were applied to the decision-making problems, one of which was based on the Smart Predict-then-Optimize (SPO) framework and named as the SPO tree algorithm. Compared with other decision tree algorithms, the SPO tree pays more attention on the "quality" of decision rather than minimizing the prediction error and provides better decision and lower model complexity. However, it remains a problem that how to apply the SPO tree to the classification task in semi-supervised learning scenario. To address such a problem, in this paper, the semi-supervised SPO tree classifier is proposed based on the density peak clustering (DPC) framework. The proposed method can utilize the information of labels, densities and distances from data. The experimental results show that, compared with other algorithms, the proposed method has a more robust classification performance in the semi-supervised learning scenario.