A Meta‐Reinforcement Learning Framework Using Deep Q‐Networks and GCNs for Graph Cluster Representation

ABSTRACT Background The rapid evolution of Internet of Things (IoT) technologies has driven innovations across domains such as robotics, autonomous systems, and environmental control. However, effectively learning graph‐based representations within these dynamic and heterogeneous systems remains a significant challenge, especially when scalability and adaptability are required. Aims This study aims to develop and evaluate a novel meta‐reinforcement learning (meta‐RL) framework that combines Deep Q‐Networks (DQNs) with Graph Convolutional Networks (GCNs) to learn adaptive and efficient representations of graph clusters. The primary objective is to enhance cluster‐based representation learning by integrating reinforcement learning with graph aggregation policies. Methods We propose a cluster policy‐GNN model that formulates optimal graph aggregation as a Markov Decision Process (MDP). The framework incorporates a cluster meta‐policy to guide node‐specific aggregation strategies and utilizes a combination of DQN and GCN for adaptive graph representation. Training involves clustering nodes based on policy‐determined hops and batching to ensure efficient GNN training. A custom reward function drives the reinforcement learning process to prioritize computational focus on the most informative subgraphs. Results Our experimental results, benchmarked on real‐world graph datasets, demonstrate that the proposed framework significantly outperforms existing state‐of‐the‐art methods, including static GNNs, alternating graph‐regularized networks, and causal‐aware neural architecture search models. The learned cluster policies effectively enhance representation learning by dynamically adjusting to the structural heterogeneity of input graphs. Improvements were observed across various domains and scales, validating the flexibility and generalizability of the method. Conclusion The proposed meta‐RL framework with integrated DQN and GCN modules offers a powerful and scalable approach for graph cluster representation learning. By introducing adaptive, node‐specific aggregation strategies guided by reinforcement learning, the method effectively captures complex graph structures and surpasses current techniques. Future work may explore real‐time adaptation and deployment in more dynamic IoT‐based applications.