Classifier Selection

We conducted a comparative experiment assessing the effectiveness of different classifiers in detecting glitch tokens using GlitchProber without post-processing. The following table showed variations in performance metrics across classifiers, revealing why we finally pick SVM.

Based on the experimental results comparing various classifiers for the GlitchProber tool without post-processing, it is evident that the SVM classifier outperforms the others in terms of the overall balance between precision, recall, and F1 score. Specifically, the SVM (with C=1 and degree=3) achieved the highest F1 score of 72.55%, higher than that obtained with other classifiers like Single Linear MLP, Double Linear MLP, Random Forest, and K-Nearest Neighbors.