Social Reinforcement Learning considers multi-agent systems with large number of agents and relatively few interactions between them, which is challenging due to high-dimensional search space, inter-agent dependencies that increase computational complexity. Moreover sparse agent interactions produce insufficient data to capture higher-order relations (interactions) for learning accurate policies. To overcome these challenges, we present a dynamic cluster-based Social RL approach that utilizes the properties of the social network structure, agent interactions, and correlations to obtain a compact model to represent network dynamics.