GT-MuRec: A Multi-Scale Graph-Theoretic GNN Framework for Dynamic Music Recommendation

Abstract

The rapid growth of online music platforms has transformed the recommendation problem into a large-scale graph problem, where millions of users and tracks create a complex network. The majority of the current graph neural network models depend on the uniform or implicitly derived weights, overlooking the explicit structural attributes of the interaction graphs. This paper introduces GT-MuRec (Graph-Theoretic Recommender), a unified framework that integrates graph-theoretic aspects with neural network message passing for dynamic and scalable music recommendation. GT-MuRec constructs a heterogeneous graph of users, tracks, singers, and albums. It redefines representation learning as a process of selective graph diffusion, where edge importance is determined by interpretable structural priors rather than uniform normalization. Experiments were conducted on chronologically split parts of the LastFM dataset using Recall@K and MRR@K under a consistent top k-ranking protocol. GT-MuRec showcases comparative early-ranking precision with an MRR@10 of 24.14% and keeps a stable retrieval performance when compared to the baseline models. GT-MuRec achieves up to 9.1% better Recall@10 and 1.4% better MRR@10 than strong GNN baseline models, and these improvements are statistically significant (p<0.01). Ablation analysis shows that each structural prior helps, especially under sparse and cold start regimes. The results show that integrating graph-theoretic ideas to neural propagation makes a music recommendation system that can grow and is structurally aware.