Heterogeneous hypergraph representation learning for link prediction

Abstract: Abstract Link prediction aims to forecast potential new links or identify and remove spurious existing links based on the current network structure. It has been successfully applied in various domains, such as gene interaction prediction in biology, criminal intelligence in social networks, co-authorship analysis in academic collaborations, and increasing sales in e-commerce by recommending product selections to customers. As informal businesses increasingly adopt e-commerce platforms, the demand for intelligent product selection mechanisms from the seller’s perspective has become more critical to ensure competitiveness and sustainability. However, using the widely used bipartite customer-product network has overlooked the seller perspective. Therefore, the study has two objectives: propose a link prediction model for seller product selection based on seller perspective and enrich the model based on domain-specific information extracted from customer feedback. To achieve this, we construct an enriched network structure by integrating customer behaviour and seller attributes into bipartite and tripartite network settings. The bipartite product-seller network is enhanced via a one-mode product projection informed by customer co-purchase behaviour. The tripartite customer-product-seller network is further extended by incorporating seller domain-specific information extracted from feedback. Extensive experiments on a real-world e-commerce dataset using Graph Convolutional Networks (GCN) are evaluated using accuracy, precision, recall, and AUC. Overall, enhanced bipartite performance has increased by 2% than conventional bipartite, while enhanced tripartite has increased by 3% than conventional tripartite. It can be concluded that both types of link prediction model and enhanced model using domain-specific have significantly increased the prediction power for seller-product selection.

TL;DR: Graph Attention Networks (GATs) as mentioned in this paper leverage masked self-attentional layers to address the shortcomings of prior methods based on graph convolutions or their approximations.