一句话:提出 CodeNoCode 方法,利用 LLM 协同生成 ML 代码与 NoCode 图形界面,提升建模透明度与包容性。
针对 NoCode 预测建模工具(如 KNIME、Orange)功能增强但业界采纳不足、非技术用户使用门槛高的问题,本文提出 CodeNoCode (CNC) 方法。该方法借助 LLM 支持代码与图形界面 (NoCode) 的协同开发,利用 ML 代码的序列化结构降低自动生成建模代码的复杂度,并将建模过程与分析表示集成到图形界面中。CNC 避免中间模型格式以提升透明度,借助 LLM 实现编程语言无关,并支持编码者与非编码者协作以增强包容性。作为立场论文,本文旨在推动 LLM 辅助的 ML 建模工具研究。
In this position paper we propose what we here call a CodeNoCode (CNC) approach to the creation and evaluation of predictive models. NoCode tools are increasingly becoming more powerful, but there is a lack of understanding about the practical/industry adoption of predictive modeling NoCode tools such as KNIME or Orange. To speed-up adoption—in particular for non-technical users— we find this a good time to investigate the LLM-supported co-development of code and GUI (NoCode) for Machine Learning (ML) tasks. ML code follows a rather sequential structure which limits the complexity of the automatic creation of modeling code and as a consequence also the representations and integrations of the modeling process and analysis within a graphical interface. CNC avoids intermediate model formats (transparency), and, in support of LLMs, can be made programming language agnostic while opening up for collaboration between coders and non-coders, in support of LLM assistants (inclusivity).