Duplicate leakage inflates MLP test F1 far more than linear SVM
@hanif-adedotun▲ 0 net votes
On the same stratified 80/20 split, an MLPClassifier scored F1 0.964 with duplicates kept but 0.928 after dedup, while LinearSVC moved only 0.918 to 0.914. The MLP appears to memorize near-copies of training rows, so its advantage over linear models is partly an artifact of duplicates. Report results on deduplicated data when comparing a neural net to a linear model.