Background/Objectives: Drug–drug interactions (DDIs) are major medication-safety concerns, and experimental testing cannot cover the expanding number of drug pairs. This systematic review evaluates graph-based machine-learning methods for DDI prediction, focusing on machine-learning architectures, data integration, interpretability, reproducibility, and clinical relevance. Methods: Following PRISMA 2020, we systematically searched major databases for studies published between January 2021 and March 2026. We included studies that applied graph-based machine-learning models, particularly graph neural networks, to predict DDIs. We compared their data sources, model designs, validation methods, predictive performance, reproducibility, and clinical relevance. Because the studies used different datasets and evaluation methods, the findings were summarized narratively rather than combined statistically. Results: Sixty studies met the eligibility criteria. Methods progressed from graph convolutional networks and graph attention networks to graph transformers, contrastive learning, multimodal fusion, and LLM-enhanced representations. We found that reported improvements in prediction performance often remained study-specific. Only three studies explicitly mentioned or addressed data leakage, whereas most reviewed studies contained no explicit leakage discussion; leakage-aware drug-disjoint, temporal, and external evaluations were also uncommon. Uncertainty calibration, computational-resource reporting, complete reproducibility materials, and independently validated explanations were also limited. Conclusions: Graph-based machine learning is promising for DDI prioritization and hypothesis generation but remains insufficient for independent clinical decision-making. Future studies should use standardized benchmarks, leakage-aware validation, calibrated uncertainty, reproducible pipelines, validated explanations, and external or prospective evaluation.