Due to the exponential growth of scientific publishing, a large-scale scientific knowledge graph (SKG) has been constructed to collect entities, concepts, and relationships across various fields. These graphs improve information discovery, literature mining, and research ecosystem decision support. SKGs’ structural intricacy, dynamic evolution, and scale make efficient and adaptive mining difficult. Traditional graph mining methods, such as rule-based extraction and static embedding models, struggle with scalability, context awareness, and dynamic graph architectures. These limits hinder navigation of the scientific literature network, knowledge discovery, and relation detection. It proposes an Adaptive Reinforcement-Learning-based Scientific Knowledge Mining framework (ARL-SKM) to address these issues. The framework optimises real-time SKG exploration and exploitation techniques using reinforcement learning agents to select nodes and edges with the best probability. ARL-SKM’s graph embedding-based reward algorithms promote novel, contextually relevant, and semantically rich knowledge patterns. This adaptive approach lets scientists learn from changing data streams. In experimental assessments on benchmark academic datasets, ARL-SKM improves link prediction, relation discovery, and emerging topic identification by 12-18% and reduces exploration cost by 15%. The adaptive technique enables the system to adapt to changing and diverse graph configurations. ARL-SKM mines scientific knowledge graphs intelligently, scalably, and adaptively. Reinforcement learning-driven adaptation opens new avenues for knowledge-driven research breakthroughs, literature retrieval, and scientific discovery.
Sinha et al. (2026) studied this question.