LLM2AutoML: Zero-Code AutoML Framework Leveraging Large Language Models

Automated Machine Learning (AutoML) aims to make machine learning accessible to non-experts by minimizing technical barriers and enabling the creation of high-performance models without extensive programming knowledge. Despite advances, existing AutoML solutions often struggle with complexity and inefficiency in handling intricate tasks. Herein, we present LLM2AutoML, a zero-code AutoML framework that leverages Large Language Models (LLMs) to generate high-performance actionable ML models without coding and with explanations. LLM2AutoML enables bidirectional human-machine alignment by interpreting user intentions expressed in natural language, converting them into executable AutoML tasks, and delivering detailed analytical reports that foster user understanding and trust. We propose the template-bounded AutoML method to ensure the LLM-generated code is highly executable, enabling fully end-to-end automation, including intention parsing, model auto-selection, and hyperparameter auto-tuning. Additionally, we incorporate techniques such as adaptive loss functions and configuration recommendations to improve efficiency and performance. Experiments on the SECOM dataset from the semiconductor manufacturing industry demonstrate that LLM2AutoML enhances the automation and usability of AutoML and LLMs, achieves superior performance, and produces high-quality analytical reports. This framework presents a novel approach to advancing the effectiveness and capabilities of both AutoML and LLMs.