PANTHER: Generative Pretraining Beyond Language for Sequential User Behavior Modeling

Description

This repository contains the implementation of PANTHER and the scripts for the experiments on credit card transaction dataset, including:

• Data preprocessing pipelines
• Pre-training on CCT behavior sequence data
• Fraud detection model training

Dataset

The raw dataset is available on Kaggle: Credit Card Transactions Dataset

Prerequisites

• Python 3.8+
• Required packages (install via pip install -r requirements.txt)

Workflow

1. Data Preprocessing

python cct_preprocess.py

Configure the raw data path in cct_preprocess.py before running.

2. Model Pretraining

Pretrain PANTHER on CCT behavior sequence data:

export CONFIG_GROUPS=credit-card-transactions
export CONFIG=tf_patternrec_v5

python train_behseq.py \
    --gin_config_file="configs/${CONFIG_GROUPS}/${CONFIG}.gin" \
    --model_save_path="./ckpt/${CONFIG_GROUPS}/${CONFIG}" \
    2>&1 | tee "logs/${CONFIG_GROUPS}/${CONFIG}.log"

3. Generate Embeddings

Extract pretrained embeddings for downstream tasks:

python inference.py \
    --gin_config_file="configs/${CONFIG_GROUPS}/${CONFIG}.gin" \
    --model_save_path="./ckpt/${CONFIG_GROUPS}/${CONFIG}"

4. Fraud Detection Model

Train the DCN fraud detection model:

python cct_fraud_detection.py