<p>Music streaming platforms frequently suffer from highly sparse user–item interactions, where many listeners provide only limited playback or skip behaviors. Such sparsity weakens the reliability of conventional similarity-based recommendation methods and often leads to unstable predictions, especially for new users, emerging music tracks, or geographically distributed service environments. To address this challenge, this study introduces an edge-assisted similarity analysis framework for robust music recommendation in sparse data settings. The framework leverages edge nodes to capture short-term listening patterns and to perform lightweight approximate similarity search, thereby reducing cloud-side computation and communication overhead. Beyond direct neighbor identification, a complementary neighbor inference mechanism is employed to exploit contrasting listening tendencies among users, enabling the system to expand candidate sets when conventional similarity evidence is insufficient. The cloud layer subsequently integrates multi-edge similarity information and refines recommendation scores under a unified optimization scheme. Experiments on a controlled sparse-interaction benchmark show that RMR achieves lower prediction error, higher recall, and competitive runtime compared with the selected baselines. These results suggest that edge-side neighbor retrieval can provide a practical way to stabilize recommendation when user histories are short and direct interaction overlap is limited.</p>