Abstract
Event log data, recording fine-grained user actions and system events,represent one of the most valuable assets for modern digital services. However,the complexity and heterogeneity of industrial event logs--characterized bylarge scale, high dimensionality, diverse data types, and intricate temporal orrelational structures--make feature engineering extremely challenging. Existingautomatic feature engineering approaches, such as AutoML or genetic methods,often suffer from limited explainability, rigid predefined operations, and pooradaptability to complicated heterogeneous data. In this paper, we propose FELA(Feature Engineering LLM Agents), a multi-agent evolutionary system thatautonomously extracts meaningful and high-performing features from complexindustrial event log data. FELA integrates the reasoning and codingcapabilities of large language models (LLMs) with an insight-guidedself-evolution paradigm. Specifically, FELA employs specialized agents--IdeaAgents, Code Agents, and Critic Agents--to collaboratively generate, validate,and implement novel feature ideas. An Evaluation Agent summarizes feedback andupdates a hierarchical knowledge base and dual-memory system to enablecontinual improvement. Moreover, FELA introduces an agentic evolutionalgorithm, combining reinforcement learning and genetic algorithm principles tobalance exploration and exploitation across the idea space. Extensiveexperiments on real industrial datasets demonstrate that FELA can generateexplainable, domain-relevant features that significantly improve modelperformance while reducing manual effort. Our results highlight the potentialof LLM-based multi-agent systems as a general framework for automated,interpretable, and adaptive feature engineering in complex real-worldenvironments.