Context-Enriched Temporal Knowledge Graph with Reinforcement Learning for Adaptive Music Recommendation in Entertainment Platforms

¹ Department of Computer Science and Engineering, HKBK College of Engineering, Bengaluru, India
² Department of Visual Communication, Kumaraguru College of Liberal Arts and science, Coimbatore, India
³ Department of Electronics and Communication Engineering, Nitte Meenakshi Institute of Technology, Nitte (Deemed to be University), Bengaluru
⁴ Megan Soft INC, Livonia, United States

^* Corresponding author: acharyamohan036@gmail.com

Abstract

Currently, Music Recommendation (MR) platforms are involved in exploiting meaning-based representations of music data to deliver personalized recommendations. These models enhance music discovery by analyzing user preferences, song attributes, and contextual data. However, the existing frameworks heavily depend on static user-item interactions and knowledge graphs, which leads to overlooking dynamic factors that reduce adaptivity in real-world entertainment platforms. Hence, this study proposes a Context-enriched Temporal Knowledge Graph with Reinforcement Learning (CTKG-RL) for adaptive MR in entertainment platforms. Initially, the input data were collected from the Last.FM_ 1 K dataset, which contains user-track interactions and then preprocessed with timestamp conversion, Levenshtein string matching, and thresholding-based filtering. This process ensures that the model has consistent time alignment and deduplication of track metadata and retains only active users and items, respectively. Further, a TKG is constructed from semantic and contextual triples, and then the embeddings are learned using Temporal Graph Attention Networks (TGANs) to capture evolving user interests. Subsequently, session modeling with a Graph Recurrent Unit (GRU)-based encoder aggregates short- and long-term preferences, and finally, an actor-critic RL model optimizes adaptive playlist generation. The proposed CTKG-RL achieved better results in terms of accuracy (0.893) when compared to the existing Multi-channel and Multi-loss MR Knowledge Graph (MM-MRKG) model.