This paper aims to address several limitations of current graph neural network (GNN)-based bundle recommender systems. These models often suffer from representation degradation as the number of aggregation layers increases and neglect latent interaction patterns that are essential for improving recommendation performance. To address these limitations, we propose a novel model named hierarchical graph-based mixup for bundle recommendation (HGMBR), which integrates mixup techniques across both the item and bundle views to learn higher-quality representations for recommendation.
The proposed model, HGMBR, uses a multilayer perceptron (MLP)-based residual network to extract latent interaction features from user–bundle and user–item embeddings. A graph-based mixup module then performs interpolation within both the bundle and item views. This process effectively enriches the user and bundle representations by capturing both explicit and latent interactions. In addition, a cross-view mixup module aligns user representations across the two views.
Extensive experiments on three publicly available data sets demonstrate that HGMBR alleviates representation degradation in deep aggregation layers and mitigates bundle sparsity issues, demonstrating its superiority over existing models in bundle recommendation tasks.
HGMBR captures latent interaction features and uses mixup techniques hierarchically to enrich user and bundle representations. This results in robust, discriminative embeddings that are less susceptible to noise and significantly enhance recommendation performance.
