Enhancing graph neural network based sequential recommender model with knowledge graph embeddings

The growth of online interactions and user-generated data has presented new challenges and opportunities for personalized recommendation systems. Traditional recommendation approaches rely on static user-item attributes and fail to account for temporal dynamics, limiting their ability to adapt to evolving preferences. Sequential recommendation models address the challenge of next-item prediction by learning from action sequences; however, their performance is limited by the lack of large, diverse, and up-to-date data. This thesis explores the integration of open-source knowledge through knowledge graphs, in the form of knowledge graph embeddings (KGEs), into graph neural network-based sequential recommendation model to enhance both semantic and sequential patterns captured within user-item interactions. By leveraging embeddings from open-linked data sources such as Wikidata, we enrich item attributes, broadening the feature set and providing a more comprehensive understanding of user preferences. Using the Movielens dataset as a foundational benchmark, four translation-based KGE models are individually incorporated within a sequential model to evaluate their impact on recommendation accuracy. Offline evaluations with Hit@K and NDCG@K metrics reveal that this approach improves the relevance and personalization of recommendations by effectively capturing both semantic context and sequential behaviors. The findings underscore the value of incorporating open-source knowledge bases into recommendation systems, demonstrating that such enriched data sources yield more accurate and meaningful recommendations.This study demonstrates how external knowledge-enhanced user and item representations improve recommendation models.