Memory-Augmented LLM-based Multi-Agent System for Automated Feature Generation on Tabular Data

论文信息¶

作者: Fengxian Dong, Zhi Zheng, Xiao Han, Wei Chen, Jingqing Ruan, Tong Xu, Yong Chen, Enhong Chen
提交日期: 2026-04-14

摘要¶

Automated feature generation extracts informative features from raw tabular data without manual intervention and is crucial for accurate, generalizable machine learning. This paper proposes a multi-agent system where different agents specialize in different feature generation strategies, coordinated by a shared memory system. The memory stores previously generated features, their performance on different datasets, and the data characteristics they were effective for. When generating features for a new dataset, agents query memory to retrieve similar historical contexts and their successful feature patterns. Different agents handle different modalities: one generates statistical aggregates, another creates interaction features, a third handles temporal patterns.

核心贡献¶

多 Agent 特征生成: 不同 Agent 专精不同特征策略（统计聚合、交互特征、时序模式）
跨数据集记忆迁移: 存储历史特征效果，检索相似数据集的成功模式
嵌入空间记忆组织: 类似数据上下文在嵌入空间聚集

为什么重要¶

展示了记忆如何实现跨数据集的知识迁移，对于数据稀缺的边缘场景特别有价值。

与端侧/移动端的相关性¶

特征生成记忆可用于端侧数据稀缺场景（如个性化健康监测），通过相似用户数据的记忆迁移弥补本地数据不足。