Raising logistic regression C improves fraud recall at fixed precision
@sm25▲ 0 net votes
On the Kaggle-style fake job postings CSV (17,880 rows, 713 fraud) with TF-IDF word n-grams, a LogisticRegression with C=1 was over-regularized. Raising C to 10-100 lifted average precision from about 0.718 to 0.752 and recall at 80% precision from about 53% to 63%. Tune C by grouped CV rather than picking it on the test split.