People and Management Related Debt in ML-Integrated Software Development Projects: Structuring Insights from Industry

The accelerated development of Machine Learning (ML) tools, combined with broader access to frameworks and infrastructures, has driven the rapid adoption of ML-based solutions in industry. However, their integration into software systems introduces unique challenges, particularly for managing technical debt (TD). While existing frameworks\/standards such as Cross-Industry Standard Process for Data Mining (CRISP-DM) and ISO\/IEC 5338 provide guidance for ML development, they fail to address the complex interplay of technical and nontechnical factors contributing to TD. Traditional TD research focuses primarily on technical issues, but in ML systems, people, and management factors, referred to as nontechnical debt (NTD), play a critical role in TD accumulation and persistence. In this study, we investigate the underexplored dimension of NTD in ML-integrated software systems, focusing on people and management-related factors. Using Design Science Research (DSR) methodology, we developed an artifact that categorizes NTD issues in ML systems. As part of this process, we conducted semi-structured interviews with 18 professionals from 15 companies, examining 22 ML projects. Through thematic analysis, we identified 15 NTD categories, each associated with underlying causes, short-term fixes, and potential solutions. Our findings indicate that NTD in ML projects often arise from decision-making practices, team dynamics, and communication barriers, all of which substantially affect project outcomes. While temporary fixes (i.e., band-aid solutions) may provide short-term relief, they frequently contribute to the accumulation over time. To support practitioners and researchers, our study complements the proposed artifact with actionable recommendations informed by expert perspectives and literature.

机器学习（Machine Learning，ML）工具的加速迭代，加之各类框架与基础设施的可及性持续提升，推动了基于ML的解决方案在工业界的快速普及。然而，将此类方案集成至软件系统时会带来独特挑战，尤以技术债务（Technical Debt，TD）的管理问题最为突出。现有诸如跨行业数据挖掘标准流程（Cross-Industry Standard Process for Data Mining，CRISP-DM）与ISO/IEC 5338等框架及标准，虽可为ML开发提供规范指导，但未能覆盖导致技术债务产生的技术与非技术因素间的复杂交互机制。传统技术债务研究多聚焦于技术层面的问题，但在ML系统中，人员与管理相关的非技术债务（Non-technical Debt，NTD）因素，对技术债务的积累与持续存续起到了关键性作用。本研究针对ML集成软件系统中尚未得到充分探索的非技术债务维度展开探究，重点关注人员与管理相关的影响因素。本研究采用设计科学研究（Design Science Research，DSR）方法论，开发了一款用于对ML系统中非技术债务问题进行分类的研究制品。在此研究过程中，我们对来自15家企业的18名行业专业人员开展了半结构化访谈，对22个ML项目进行了系统性分析。通过主题分析方法，我们共识别出15类非技术债务类别，每一类均对应其潜在成因、短期修复手段与潜在解决方案。研究结果显示，ML项目中的非技术债务往往源于决策实践不当、团队动态失衡与沟通壁垒，这些因素均会对项目产出造成显著影响。尽管权宜之计（即创可贴式解决方案）可提供短期缓解，但此类手段往往会随着时间推移加剧技术债务的积累。为赋能从业者与研究者，本研究结合专家调研观点与现有学术文献，为所提出的研究制品补充了具备可操作性的实践建议。