.png){kind=icon}
Return
ProcessML: A Process-Aware Agentic Framework for Service-Oriented Machine Learning Pipeline Generation and Optimization
.png){kind=icon}
PRE
AI
DOI：10.1109/tsc.2026.3731106.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Automated generation and optimization of machine learning (ML) pipelines as services can substantially reduce the workload of ML engineers while democratizing ML for non-experts. While Large Language Models (LLMs) offer strong knowledge and coding capabilities, a single LLM often struggles to handle end-to-end pipeline design, especially for increasingly complex multimodal and multidomain ML tasks. As a result, LLM-based agentic frameworks that leverage multi-agent collaboration have emerged as a more effective and scalable solution. However, existing approaches often fail to capture interdependencies between pipeline components or lack fine-grained feedback mechanisms to support iterative refinement. To address these limitations, we propose ProcessML, a process-aware agentic framework for service-oriented machine learning pipeline generation and optimization, in which ML pipelines are constructed, executed, and iteratively refined within a service-oriented lifecycle. ProcessML incorporates two key process-aware mechanisms: Contextual Dependency-Based Pipeline Modeling, which explicitly captures and models intra-process dependencies during pipeline generation, and Empirical Feedback-Driven Iterative Refinement, which enables systematic learning from past iterations for fine-grained pipeline optimization. Together, these mechanisms enable contextual consistency, continuous improvement, and more adaptive ML pipeline generation. We implement ProcessML as a service-oriented prototype and evaluate it on a state-of-the-art real-world benchmark. Extensive experimental results demonstrate that ProcessML consistently outperforms existing frameworks, with further analyses confirming its effectiveness and robustness in ML pipeline generation and optimization.
Keywords:
Large Language Models
Agentic Framework
Service-Oriented
AutoML
Process-Aware Mechanisms
Journal
IF:
5.8
Papers:
2.2K
Citations:
6.5K
Organization
View more organizations
Cited Papers
No cited papers available