Pathways, evidence types, and interpretive boundaries in machine learning-assisted source analysis of soil metal(loid)s.

Machine learning (ML) is increasingly incorporated into source-related analytical tasks for soil metal(loid)s, including source structure identification, quantitative source apportionment, and environmental driver analysis and model interpretation. However, existing reviews lack a systematic framework organized around environmental management tasks and evidence types. We systematically searched Web of Science and Scopus following PRISMA and included 204 studies. Bibliometric analysis showed rapid growth since 2020, with supervised learning more common than unsupervised learning and methodological integration becoming a stable pattern. We propose three representative pathways: source structure identification and pattern characterization, quantitative source apportionment and model coupling, and environmental driver analysis and model interpretation. These correspond to qualitative structural evidence, quantitative source contribution evidence, and semi-quantitative explanatory evidence, respectively, addressing how pollution patterns are organized, how much different sources contribute, and how environmental variables are associated with pollution patterns or source-related spatial variation. We further compare their interpretability and evidence functions and propose a pathway selection strategy for different management objectives. This review guides selection of appropriate ML pathways, reduces evidence misinterpretation, and provides decision support for soil pollution management.