AI CUP 2023 玉山人工智慧公開挑戰賽－信用卡冒用偵測

This method achieved 7th position on the private leaderboard (TEAM_4043).

Final competition slide [link]

There is no new idea in this method. Just calculated some rule-based features (basic statistics) and classified them by XGBoost classifier.

Note: I have not optimized the process. The processing steps are very time-consuming and consume a lot of memory (>50GB). To proceed, ensure you have a good CPU, GPU, and sufficient RAM.

The preprocessing data and checkpoints are not available now. If you need the checkpoints, please contact me.

• Preprocessing table: ~14 GB
• XGBoost models: ~400 MB per model

The data preprocessing process involves concatenating the training, public_test, and private tables to calculate basic statistics for the "cano" and "chid" groups.

Note: Before the preprocessing, you should place the tables in the tables directory. (The filename of these tables should be training.csv, public.csv and private.csv).

cd Preprocessing
python preprocessing.py -o output/preprocessing.csv

Note: This step could take several hours to complete.

The model is XGBoost. The file model.py contains the parameters which are currently unchangeable.

python train.py \
    --input output/preprocessing.csv \
    --model_output_dir  output/checkpoints/ \
    --thr_path config/your_thr.json \
    --epochs 300 \
    --runs 3 \ --> Number of models (for ensemble)
    --gpu 0

Note: This step could take about 1 hour to complete by using GPU.

To perform inference of the data without preprocessing and training your model, download the preprocessing table and model checkpoints first. Then, move them to the output directory.

python inference_submit.py \
    --input output/preprocessing.csv \
    --thrs config/thr.json \ --> Best thresholds of my models
    --ckpts output/checkpoints/ \
    --output submission.csv

Note: After inference, you must merge the "txkey" of the example submission file to get the correct submission.

All the code can be found in the final_code directory

Step 1: Preprocessing. Use training.csv provided before the final competition, private_1.csv and private_2_processed.csv to calculate basic statistics for the "cano" and "chid" groups as features. These files should be placed in the tables directory.

However, the CPU performance is insufficient to complete all the preprocessing. So we only transform some essential features (Only including the transformed features of numerical features).

cd final_code

python preprocess_numerical.py -o ../output/preprocessing_final.csv

Step2: Since the environment of the final competition is no GPU, we change the device of XGBoost classifier to "cpu". Also, you must first turn off the parameters subsample and sampling_method in model.py because they are only available on GPU.

python train_numerical.py \
    --input ../output/preprocessing_final.csv \
    --model_output_dir  ../output/checkpoints/ \
    --thr_path thr_final.json \
    --epochs 100 \
    --runs 3 \
    --gpu cpu

Step3: Inference (ensemble 3 models)

python inference_numerical.py \
    --input ../output/preprocessing_final.csv \
    --thrs thr_final.json \
    --ckpts ../output/checkpoints/ \
    --output submission.csv

Note: After step 3, you must merge the "txkey" of the example submission file to get the correct submission.