Application of Classification and Regression Tree and Spectral Clustering to Breast Cancer Prediction: Optimizing the Precision-Recall Trade-Off

The Receiver Operator Characteristic (ROC) test is often used to evaluate classification performance. However, it calls for special consideration when applied to the class-imbalanced data. The Precision-Recall Curve (PRC) is dependent on the class imbalance ratio. It summarizes the trade-off between the true positive rate and the positive predictive value for a predictive model using different probability thresholds. In this work, PRC is used to enhance the precision and recall of the Classification And Regression Tree (CART) algorithm to diagnose breast cancer; this enhanced the CART recall’s capacity to predict wrongly classified samples by 10%, from 89% to 99%, when compared to the standard CART algorithm. On the other hand, when evaluating the performance of the improved CART on unseen data, it achieved a recall of 95%. Thus, the recall of the CART decreased by 4% from 99% to 95%. This drawback can be attributed to the small size of the training data, which has been labeled using one of the clustering methods, spectral clustering, which achieved the best Silhouette in separating data into two clustering with a score of 0.43.