jAEhEEkIM/operationbench-toy-drift-streams

该数据集是一个为OperationBench设计的经过处理的玩具数据流包，专门用于代理概念漂移实验，其中环境在每个时间点发出强制周期性观察包。它包含三个来源数据流：electricity作为渐变漂移基线，noaa_weather作为循环漂移基线，credit_card_fraud作为突变和标签偏移压力测试。数据集采用共享schema，包括来源信息、漂移配置、流索引、特征列（最多30个）和标签等字段。注意：electricity存在标签依赖性和捷径行为，credit_card_fraud极度不平衡，NOAA Weather缺乏明确的再分发许可，需谨慎使用。This dataset is a processed toy data stream bundle for OperationBench. It is designed for agentic concept-drift experiments where the environment emits a forced periodic observation packet at every tick. It includes three source streams: electricity as a gradual drift baseline, noaa_weather as a recurring drift baseline, and credit_card_fraud as an abrupt and label-shift stress test. The dataset uses a shared schema with fields such as source information, drift profile, stream indices, features (up to 30 columns), and labels. Note: electricity has label dependency and shortcut behavior, credit_card_fraud is extremely imbalanced, and noaa_weather lacks a modern explicit redistribution license, requiring careful usage.