Precision–recall curve (PRC) classification trees

The classification of imbalanced data has presented a significant challenge for most well-known algorithms that were often designed with relatively balanced class distributions. Nevertheless skewed distribution is common feature in real world problems. It especially prevalent certain application domains great need machine learning and better predictive analysis such as disease diagnosis, fraud detection, bankruptcy prediction, suspect identification. In this paper, we propose novel tree-based algorithm based on the area under precision-recall curve (AUPRC) variable selection context. Our algorithm, named "Precision-Recall Curve tree", or simply "PRC tree" modifies two crucial stages tree building. first stage to maximize node selection. second harmonic mean recall precision (F-measure) threshold We found proposed PRC tree, its subsequent extension, random forest, work well class-imbalanced sets. have demonstrated our methods outperform their classic counterparts, usual CART forest both synthetic data. Furthermore, ROC by group previously shown good performance combination them, PRC-ROC also shows promise identifying minority class.