WARMER: Meta‐Learned Graph Warming for User Cold‐Start Recommendation

Abstract

The user cold‐start recommendation problem is one of the critical challenges for recommender systems, making it difficult to deliver precise and tailored recommendations. Previous research often attempts to enhance representations through indirect connections within bipartite graph structures or Heterogeneous Information Networks. However, these methods fail to fully capitalize on direct user‐item interaction patterns. To address this challenge, we propose WARMER, a novel meta‐learned graph warming framework for user cold‐start recommendation. WARMER leverages pseudo‐labelling on cold‐start interactions, along with meta‐learning strategies, to perform robustly even with minimal initial data. Specifically, the model initialized with meta‐learned parameters adapts to user‐specific contexts from a limited support set and generates warming edges to enrich the sparse interaction graphs. WARMER comprises a ‘Warming’ network, which transforms a cold graph into a warm graph by predicting warming edges among unlabelled data, and an ‘Estimator’ network that utilizes these predicted edges to learn the user ratings. Extensive experiments demonstrate that WARMER effectively bridges the gap between cold and warm graph characteristics, achieving state‐of‐the‐art performance in cold‐start recommendations.