Master of Science (MS)
RelationalNet, Graph Neural Networks, Recommendation Systems
Traditional recommender systems create models that can predict user interests based on the user-item relationships. However, these systems often have limited performance due to sparse user behavior data. To address this challenge, researchers are now exploring models for social recommendation that can account for both user- user and user-item relationships based on social networks, and user past behavior, respectively. These models aim to understand each user’s behavior by considering their trusted neighbors and their influence on each other. Specifically, the potential embedding of each user is influenced by their trusted neighbors, who are, in turn, influenced by their own trusted neighbors and social connections. Users’ interests evolve over time as social influence propagates recursively in the social network.
In this project, we propose the RelationalNet model, which creates graphs not only for the user-item and user-user relationships but also for the item-item relationships. We learn the user interest predictions by using Graph Neural Networks(GNNs). By incorporating social influence into recommendation models, we can capture more accurate user interests, especially when traditional methods fall short due to data sparsity. Such models improve the accuracy and effectiveness of recommendation systems by leveraging social connections and interactions. Moreover, we perform ex- periments that demonstrate that incorporating the item graph information into GNNs produces better results compared to the current cutting-edge social recommendation models.
Tallapally, Dharahas, "RelationalNet using Graph Neural Networks for Social Recommendations" (2023). Master's Projects. 1230.
Available for download on Friday, May 24, 2024