基于大小模型协同优化的表格数据准备工作流生成方法

doi:10.3969/j.issn.1000-5641.2026.05.008

华东师范大学学报（自然科学版） ›› 2026, Vol. 2026 ›› Issue (5): 95-108.doi: 10.3969/j.issn.1000-5641.2026.05.008

基于大小模型协同优化的表格数据准备工作流生成方法

陈思蓓¹^,²(), 韩宵玥², 范举²^,³^,*(), 杜小勇²^,³

1. 重庆大学计算机学院, 重庆　401331
2. 中国人民大学信息学院, 北京　100872
3. 中国人民大学数据工程与知识工程教育部重点实验室, 北京　100872

收稿日期:2026-07-08 出版日期:2026-09-25 发布日期:2026-09-12
通讯作者: 范举 E-mail:chen.sibei@cqu.edu.cn;fanj@ruc.edu.cn
作者简介:陈思蓓, 女, 助理研究员, 研究方向为智能数据治理. E-mail: chen.sibei@cqu.edu.cn
基金资助:
国家自然科学基金 (62436010, 62441230); 教育部中央高校青年教师科研创新能力支持项目 (RICSPYF-ZY2025001)

Collaborative optimization of large and small models for tabular data-preparation pipeline generation

Sibei CHEN¹^,²(), Xiaoyue HAN², Ju FAN²^,³^,*(), Xiaoyong DU²^,³

1. College of Computer Science, Chongqing University, Chongqing　401331, China
2. School of Information, Renmin University of China, Beijing　100872, China
3. Key Laboratory of Data Engineering and Knowledge Engineering (Ministry of Education), Renmin University of China, Beijing　100872, China

Received:2026-07-08 Online:2026-09-25 Published:2026-09-12
Contact: Ju FAN E-mail:chen.sibei@cqu.edu.cn;fanj@ruc.edu.cn

摘要/Abstract

摘要：

随着大数据与人工智能的发展, 越来越多的行业用户开始借助机器学习进行数据分析. 本文聚焦于机器学习场景下的数据准备问题, 旨在针对给定的机器学习任务 (包括表格数据集、训练目标和模型配置), 自动构建高质量数据准备工作流. 现有工作通常基于深度神经网络模型, 这类方法需要预先定义一个有限的数据准备操作空间, 难以满足定制化需求. 随着大语言模型 (LLM) 的发展, 越来越多的研究开始探索利用 LLM 自动生成数据准备工作流, 但由于 LLM 难以理解表格数据的结构, 并且缺乏面向下游任务性能的目标优化, 限制了模型的最终表现. 为此, 本文提出了一种新颖的基于大小模型协同优化的表格数据准备工作流生成框架——LLMPipe. 该方法首先充分利用大语言模型在语义理解与知识泛化方面的优势, 结合任务描述与上下文信息, 生成具有针对性和领域适应性的定制化数据准备操作; 随后, 引入轻量级模型对操作特征进行建模, 并利用强化学习算法优化操作排序策略, 从而实现对数据准备流程的动态编排. 该框架通过“大模型扩充、小模型调优”的协同方式, 在保障领域知识理解能力与高质量工作流组合能力的同时, 提高了整体计算效率, 有效增强了生成工作流的定制性与合理性. 本文在 Kaggle 平台上的多个真实机器学习任务上开展了大量实验. 实验结果表明, 相较于现有方法, 本文提出的框架在机器学习任务上的准确性取得了显著提升, 达到了当前最优水平.

关键词: 机器学习, 数据准备工作流, 强化学习, 大语言模型, 自动机器学习

Abstract:

Owing to the development of big data and artificial intelligence, industry users are increasingly leveraging machine learning for data analysis. This study focuses on data preparation for machine-learning scenarios to automatically construct a high-quality data-preparation pipeline for a specified machine-learning task (including a tabular dataset, training objective, and model configuration). Existing methods typically rely on deep neural networks and require a predefined and limited operation space for data preparation, making it difficult to satisfy customization demands. Researchers are increasingly investigating the use of large language models (LLMs) to automatically generate data-preparation pipelines. Nevertheless, LLM have difficulty understanding the structure of tabular data and lack performance-oriented optimization for downstream tasks. This ultimately limits model performance. Hence, this study proposes a novel framework named LLMPipe to generate a tabular-data preparation pipeline based on the optimization of large and small models. The method first leverages the advantages of LLM in semantic understanding and knowledge generalization to generate customized and domain-adaptive data-preparation operations by incorporating task descriptions and contextual information. Subsequently, it introduces a lightweight model to model operation features and uses reinforcement learning to optimize the operation ordering strategy, thereby dynamically orchestrating the entire data-preparation pipeline. Based on a collaborative “large-model expansion and small-model tuning” approach, the framework preserves the ability to understand domain knowledge and compose high-quality pipelines while improving overall computational efficiency and effectively enhancing the customizability and rationality of the generated pipelines. Extensive experiments are conducted across multiple real-world machine-learning tasks from Kaggle. The results reveal that the proposed method significantly outperforms existing methods, with state-of-the-art accuracy achieved across machine-learning tasks.

Key words: machine learning, data preparation pipeline, reinforcement learning, large language model, automatic machine learning

中图分类号:

TP18

陈思蓓, 韩宵玥, 范举, 杜小勇. 基于大小模型协同优化的表格数据准备工作流生成方法[J]. 华东师范大学学报（自然科学版）, 2026, 2026(5): 95-108.

Sibei CHEN, Xiaoyue HAN, Ju FAN, Xiaoyong DU. Collaborative optimization of large and small models for tabular data-preparation pipeline generation[J]. J* E* C* N* U* N* S*, 2026, 2026(5): 95-108.

图/表 7

图1

图2

图3

图4

表1

表2

表3

参考文献 16

1	Chen S B, Tang N, Fan J, et al.. HAIPipe: combining human-generated and machine-generated pipelines for data preparation. Proceedings of the ACM on Management of Data, 2023, 1 (1): 1- 26.
2	Heffetz Y, Vainshtein R, Katz G, et al. DeepLine: AutoML tool for pipelines generation using deep reinforcement learning and hierarchical actions filtering [C]//Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. ACM, 2020: 2103-2113.
3	Li P, Chen Z Y, Chu X, et al.. DiffPrep: differentiable data preprocessing pipeline search for learning over tabular data. Proceedings of the ACM on Management of Data, 2023, 1 (2): 1- 26.
4	Feurer M, Klein A, Eggensperger K, et al. Efficient and robust automated machine learning [C]//Advances in Neural Information Processing Systems 28. Curran Associates, Inc, 2015: 2962-2970.
5	Olson R S, Moore J H. TPOT: a tree-based pipeline optimization tool for automating machine learning [C]//Proceedings of the 2016 Workshop on Automatic Machine Learning. PMLR, 2016: 66-74.
6	Hilprecht B, Hammacher C, Reis E S, et al. DiffML: end-to-end differentiable ML pipelines [C]//Proceedings of the Seventh Workshop on Data Management for End-to-End Machine Learning. ACM, 2023: 1-7.
7	Kingma D P, Ba J. Adam: a method for stochastic optimization [PP/OL]. V9. arXiv (2017-01-30)[2026-07-10]. https://arxiv.org/abs/1412.6980.
8	Mnih V, Kavukcuoglu K, Silver D, et al. Playing Atari with deep reinforcement learning [PP/OL]. V1. arXiv (2013-12-19)[2026-07-10]. https://arxiv.org/abs/1312.5602.
9	Trirat P, Jeong W, Hwang S J. AutoML-agent: a multi-agent LLM framework for full-pipeline AutoML [C]//Proceedings of the 42nd International Conference on Machine Learning. PMLR, 2025, 267: 60099-60146.
10	Pedregosa F, Varoquaux G, Gramfort A, et al.. Scikit-learn: machine learning in Python. The Journal of Machine Learning Research, 2011, 12, 2825- 2830.
11	Luo D Q, Feng C J, Nong Y X, et al. AutoM³L: an automated multimodal machine learning framework with large language models [C]//Proceedings of the 32nd ACM International Conference on Multimedia. ACM, 2024: 8586-8594.
12	Hong S R, Lin Y Z, Liu B, et al. Data interpreter: an LLM agent for data science [C]//Association for Computational Linguistics. ACL, 2025: 19796-19821.
13	Guo S Y, Deng C, Wen Y, et al. DS-Agent: automated data science by empowering large language models with case-based reasoning [C]//Proceedings of the 41st International Conference on Machine Learning. PMLR, 2024: 16813-16848.
14	Chi Y Z, Lin Y Z, Hong S R, et al. SELA: tree-search enhanced LLM agents for automated machine learning [PP/OL]. V1. arXiv (2024-10-22)[2026-07-10]. https://arxiv.org/abs/2410.17238.
15	Xu J L, Liu Z, Suryanarayanan N A V, et al.. Large language models synergize with automated machine learning. Transactions on Machine Learning Research, 2024, 1- 49.
16	Chen S B, Fan J, Wu B, et al.. Automatic database configuration debugging using retrieval-augmented language models. Proceedings of the ACM on Management of Data, 2025, 3 (1): 1- 27.

模型	DTC的性能		RFC的性能		KNN的性能		LR的性能
模型	准确率	成功率	准确率	成功率	准确率	成功率	准确率	成功率
Qwen3.5-Coder-Plus (480 B)	0.757	1	0.798	1	0.726	1	0.753	1
Claude-Opus-4 (300~500 B)	0.788	1	0.834	1	0.781	1	0.809	1
Qwen2.5-72B-Instruct (72 B)	0.670	0.820	0.717	0.836	0.598	0.754	0.555	0.672
LLMPipe (72 B)	0.804	1	0.854	1	0.823	1	0.824	1

方法/模型	准确率	成功率
Qwen3.5-Coder-Plus (480 B)	0.758	0.918
Claude-Opus-4 (300~500 B)	0.788	0.967
Qwen2.5-72B-Instruct (72 B)	0.631	0.738
AutoML-Agent (72 B)	0.275	0.311
AutoSklearn	0.803	1
DiffPrep	0.758	1
LLMPipe (72 B)	0.863	1

模型	准确率
模型	DTC	RFC	KNN	LR
LLMPipe (LLM with full context)	0.795	0.836	0.810	0.813
LLMPipe (w/o customized operation)	0.792	0.822	0.795	0.812
LLMPipe (Random planning)	0.713	0.738	0.705	0.710
LLMPipe (LLM planning)	0.621	0.628	0.630	0.559
LLMPipe	0.804	0.854	0.823	0.824

[1]	赵乙茗, 刘昊阳, 董宜滔, 毛嘉莉, 李忠斌. 钢铁物流语义感知与不确定性量化协同调度框架[J]. 华东师范大学学报（自然科学版）, 2026, 2026(5): 133-144.
[2]	纪坤华, 金浩阳, 刘文辉, 黄定江. 大语言模型与多智能体赋能新型电力系统[J]. 华东师范大学学报（自然科学版）, 2026, 2026(5): 178-195.
[3]	张远来, 汪庆宁, 高明. SQL4Domain: 面向电力垂类领域的高可靠Text-to-SQL生成与校验框架[J]. 华东师范大学学报（自然科学版）, 2026, 2026(5): 196-207.
[4]	钱卫宁, 王春阳, 兰韵诗, 王伟, 蒲鹏, 周傲英. 语言教育科技问题、方法与实践[J]. 华东师范大学学报（自然科学版）, 2026, 2026(5): 224-236.
[5]	马毅鸣, 林欣, 倪琴, 余杨泽, 邓赐平, 贺樑. 针对大语言模型的社会认知评测框架和问答基准[J]. 华东师范大学学报（自然科学版）, 2026, 2026(4): 102-111.
[6]	宫泽正, 沈富可, 魏同权. 基于大语言模型的测试用例图谱生成研究[J]. 华东师范大学学报（自然科学版）, 2026, 2026(4): 143-153.
[7]	鲁兴宇, 吴莹, 樊俊宁, 叶雨萍, 王佳. 基于机器学习的长江溶解态有机碳输送的长期变化及驱动因素分析[J]. 华东师范大学学报（自然科学版）, 2026, 2026(3): 17-29, 43.
[8]	谢林娜, 陆雪松. 基于学生开源社区行为的数字岗位就业预测[J]. 华东师范大学学报（自然科学版）, 2025, 2025(5): 87-98.
[9]	许立君, 杨丽, 黄子祎. 大语言模型与开源生态协同的人工智能课程创新探索与研究[J]. 华东师范大学学报（自然科学版）, 2025, 2025(5): 66-75.
[10]	庞瑞洋, 陆雪松. 基于智能体的可交互数据结构和算法可视化实现[J]. 华东师范大学学报（自然科学版）, 2025, 2025(5): 32-42.
[11]	何德鑫, 韩凡宇, 王伟. 大语言模型在开源项目主题标注中的应用与评估研究[J]. 华东师范大学学报（自然科学版）, 2025, 2025(5): 14-24.
[12]	傅宇飞, 徐泽松, 冯安昊, 魏同权. 基于流量感知的移动网络授权信息在线预测方法[J]. 华东师范大学学报（自然科学版）, 2025, 2025(4): 1-14.
[13]	陈郅睿, 陆雪松. 基于开源代码大语言模型提示的学生代码修复[J]. 华东师范大学学报（自然科学版）, 2024, 2024(5): 93-103.
[14]	寇思佳, 闫凤云, 马晶. 国内大语言模型在学科知识图谱自动标注上的应用——以道德与法治和数学学科为例[J]. 华东师范大学学报（自然科学版）, 2024, 2024(5): 81-92.
[15]	刘佳, 孙新, 张宇晴. 知识图谱与大语言模型协同的教育资源内容审查[J]. 华东师范大学学报（自然科学版）, 2024, 2024(5): 57-69.

基于大小模型协同优化的表格数据准备工作流生成方法

Collaborative optimization of large and small models for tabular data-preparation pipeline generation

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 7

参考文献 16

相关文章 15

编辑推荐

Metrics

本文评价