Drug–drug interactions are a major source of preventable medication-related harm, particularly among patients exposed to polypharmacy, and machine learning has increasingly been proposed to move beyond static interaction tables by leveraging molecular, clinical, pharmacovigilance, and knowledge-graph data. This systematic review evaluated machine learning models for drug–drug interaction prediction published between 2017 and 2025, focusing on data sources, validation designs, interpretability, and evidence of clinical utility. Following a PRISMA 2020-compliant search across PubMed, Scopus, Web of Science, and IEEE Xplore, two reviewers screened records, extracted study characteristics, and synthesized the evidence narratively due to heterogeneity that precluded meta-analysis. The literature expanded substantially during this period, with many studies employing deep learning, graph neural networks, similarity-based methods, and ensemble approaches; however, model development was usually retrospective, validation predominantly internal, and direct evidence of clinical utility remained limited. Overall, machine learning shows promise for identifying potential drug–drug interactions and prioritizing clinically important risks, but before widespread implementation, the field requires stronger external validation, prospective clinical evaluation, and transparent reporting of deployment-relevant outcomes.