![]() | Only 14 pages are availabe for public view |
Abstract In the realm of modern digital applications, recommender systems have emerged as vi- tal tools, assisting users in navigating a plethora of choices. This research addresses the challenge of enhancing the accuracy of recommendations by harnessing the capa- bilities of Graph Neural Networks (GNNs) and multimodal data integration. GNNs are identified as effective tools for enriching node feature representations by propagat- ing local information. However, their efficacy decreases when dealing with data-scarce nodes. In response, this thesis introduces a novel multimodal graph-based recommen- dation system. Our approach synergises GNNs with LLMs and diverse modal features. The fusion of multiple modalities, such as text and images, improves node embeddings. This is achieved through advanced feature engineering techniques. As a result, there is a substantial improvement in recommendation precision. The evaluation, conducted on the MovieLens and Goodreads datasets, validates to the system’s efficacy in harnessing varied data sources. The concept of a unified node representation further supports the system’s capability to capture latent semantic relationships between nodes. This integra- tion is instrumental in enabling content-based recommendations, showcasing the model’s competence in encapsulating user-item interactions and rendering accurate item rating predictions. Ultimately, this research sheds light on the synergy between graph-based machine learning and recommendation systems, accentuating the pivotal role of data modalities in enhancing accuracy and personalization. The proposed approach holds the potential for fostering robust recommendation systems across diverse domains. |