Helping GNN-based Recommendation Systems Work Better When User Data Is Sparse

What is it about?

In location-based recommender systems, a user creates a connection with a location when they check in there. These interactions are often modeled as a bipartite graph, with users and locations as nodes and check-ins as edges. However, such graphs tend to be highly sparse, especially for new users who have very few check-ins. This leads to the formation of disconnected subgraphs, which prevents traditional Graph Neural Networks (GNNs) from effectively spreading information across the entire network. To solve this problem, we propose SEP-GCN, a novel model that goes beyond direct check-in connections. By using the time and geographical coordinates of check-ins, we identify hidden relationships between similar user-location interactions, even if they belong to separate parts of the graph. These new connections allow our model to share information globally, improving recommendation accuracy, especially in sparse or cold-start scenarios.

Featured Image

Photo by Jilbert Ebrahimi on Unsplash

Photo by Jilbert Ebrahimi on Unsplash

Why is it important?

Experiments show that SEP-GCN consistently outperforms other state-of-the-art GNN-based recommender systems in terms of accuracy, recall, precision, and NDCG. Additionally, we conducted evaluations under varying levels of data sparsity and found that SEP-GCN performs especially well when data is extremely sparse, making it a highly robust and practical solution for real-world recommendation tasks.