GraSS: A graph-based skip-gram synthesizer for session data augmentation

Abstract

Session-based recommender systems using graph neural networks (GNNs) have achieved strong performance by modeling item transitions within user sessions. Despite their success, these models often struggle to generalize well to infrequent or uncommon behavior patterns. Such patterns, due to their limited representation in the training data, are typically overshadowed by frequent transitions, leading to biased recommendations and reduced performance in real-world scenarios where long-tail behaviors are prevalent. To address this limitation, we introduce GraSS (graph-based skip-gram synthesizer), a novel data augmentation framework designed to improve the robustness of GNN-based session recommenders. GraSS identifies sessions containing rare item transitions and enriches them by generating synthetic session sequences. This is achieved by utilising skip-gram statistics to capture contextual item co-occurrences and applying random walks on an item graph to generate plausible but diverse session paths. The augmented sessions are then used to retrain the model, enabling better learning from sparse behaviors. Experiments on standard session-based recommendation benchmarks demonstrate that GraSS consistently improves recommendation accuracy.