gusdelact/credit-card-fraud-curated

这是一个经过整理的信用卡欺诈检测数据集，基于原始数据集alenc123/credit-card-fraud，应用了特征工程并进行了分层划分的训练/测试分割，适用于二分类欺诈检测的基准测试。数据集包含1,037,340行训练数据和259,335行测试数据，具有30个最终特征，欺诈率约为0.579%（在两个分割中均分层）。目标变量为is_fraud，取值为0或1。预处理包括删除无信号列、创建衍生特征（如小时、星期几、月份、年龄、距离等）、频率编码和独热编码。数据集文件以Parquet格式提供。需要注意的是，原始数据是合成的（基于Sparkov家族），因此性能指标可能比真实欺诈数据更乐观；频率编码基于训练集固定频率；且没有时间分割，对于概念漂移场景建议重新按日期分区。

A curated version of the credit-card-fraud-detector dataset based on the original alenc123/credit-card-fraud dataset, with applied feature engineering and stratified train/test splits. Ready for binary classifier benchmarking on fraud detection. The dataset consists of 1,037,340 training rows and 259,335 test rows, with 30 final features and a fraud rate of approximately 0.579% (stratified in both splits). The target variable is is_fraud ∈ {0, 1}. Preprocessing includes dropping non-signal columns, creating derived features (e.g., hour, dayofweek, month, age, distance_km, amt_log1p), frequency encoding, and one-hot encoding. Files are provided in Parquet format. Limitations: the original dataset is synthetic (Sparkov family), so metrics may be optimistic compared to real fraud; frequency encoding uses training split frequencies; and there is no temporal split, so re-partitioning by date is recommended for concept drift scenarios.