(A) node embedding enhancement framework to mitigate cold start problem in GNN

Abstract

The cold start problem is a significant challenge in Graph Neural Network (GNN)-based models for tasks like recommender systems and link prediction. Limited interaction nodes, known as "cold nodes", hinder accurate embedding formation, leading to performance degradation. Existing approaches often rely on complex calculations or separate side information, which may not be practical in real-world scenarios. To address this, we propose a Node Embedding Enhancement Framework (NEEF), a framework that focuses on improving node representations to mitigate the cold start problem. Inspired curriculum learning, framework generates reliable node embeddings from a subgraph of "warm nodes". These embeddings are then integrated into the graph's node features, improving discriminate power. Extensive experiments on real-world datasets demonstrate the effectiveness of our framework in mitigating cold start effects. It consistently outperforms the state-of-the-art methods, significantly improving performance. The framework seamlessly integrates into existing GNN architectures, enabling broad application without major modifications. Our improved embedding framework advances GNN-based models, addressing the cold start problem and enhancing their capabilities. Its practicality and effectiveness have the potential to enhance real-world applications that rely on graph-based data without adding side information.