Movie recommendation system integrating visual poster features and temporal dynamics

Movie recommendation systems help users discover relevant content in today's vast digital libraries. However, traditional systems often overlook two key factors: how user preferences evolve over time and the visual information conveyed by movie posters. This study proposes a comparison framework based on graph neural networks to systematically investigate how temporal information and visual features contribute to rating prediction, both individually and in combination. The study converts the MovieLens dataset, which contains user ratings, movie metadata, and poster images, converting into a dynamic heterogeneous graph with users, movies, genres, and tags as nodes. Using PyTorch Geometric, an efficient data pipeline is built that provides unified graph-structured input for the models. Four modeling approaches are then proposed and evaluated: Baseline GNN, Temporal Graph Network (TGN), Multimodal GNN (MM-GNN), Temporal-Multimodal GNN (T-MMGNN). All models were evaluated in the training set, validation set and test set that were divided in chronological order. These sets were also divided in chronological order, thus simulating a situation similar to the actual system deployment. Among the evaluated methods, the hybrid T-MMGNN model obtains the best performance, which improves 8.3% in Precision@K over the baseline. This result highlight that the joint utilization of temporal dynamics and visual information significantly improves recommendation accuracy. Furthermore, the study provides both valuable actionable insights and a reproducible framework to facilitate the development of future recommendation systems incorporating content and temporal patterns.