BAFS Benchmark Datasets: synthetic and fraud_mixed (IEEE TKDE)

Author-generated benchmark datasets accompanying the paper "Bandit-Guided Adaptive Feature Screening (BAFS)" (IEEE TKDE). Two datasets are included: - synthetic: 80,000 train / 20,000 test rows, 500 numeric features, ~30.3% positive. No header; the label is the last (501st) column. - fraud_mixed: 2,000,000 train / 500,000 test rows, ~3.37% fraud. 300 dense + 125 partially-sparse + 75 sparse features; NaN denotes missing. No header; the label is the last column. Packaging: this record is a single archive, datasets_bundle.tar.gz, containing the four CSV files plus SHA256SUMS.uncompressed and DATASET_README.md. After download: tar -xzf datasets_bundle.tar.gz sha256sum -c SHA256SUMS.uncompressed # verify the extracted CSVs The SHA256SUMS.uncompressed hashes match data/manifest.sha256 in the paper's code artifact. Third-party datasets used in the paper (UCI, Kaggle, Criteo) are NOT included here for licensing reasons; they are fetched via the code artifact's data/download.py.