The prediction of adverse drug reactions (ADRs) remains a critical challenge in pharmacovigilance due to the massive, heterogeneous, and distributed nature of biomedical and clinical data. This study proposes a big data–driven federated knowledge graph for ADR prediction that integrates distributed, heterogeneous pharmacovigilance sources, including biomedical reporting systems, electronic health records, and biomedical literature, without centralizing sensitive data. The framework utilizes scalable data management pipelines, built on Apache Spark and NoSQL databases, to handle voluminous, unstructured data. Knowledge graphs are constructed using ontology-based annotation to semantically link drugs, reactions, and patient attributes, while large language models (LLMs) are leveraged for automated knowledge extraction, entity disambiguation, and natural language querying. Graph neural networks (GNNs) are employed for multi-relational ADR risk prediction and analysis. The proposed architecture demonstrates integration of data-driven knowledge representation and federated learning to enable scalable ADR prediction systems. Across Neonatal, OFFSIDES, and ADReCS benchmarks, the proposed federated KG + LLM approach improves relation-wise prediction gain on F1-score by \(\approx \) 8–10 % and PR–AUC by 4–11% over the best baseline, while maintaining strong ROC–AUC.