Item Relationship Graph Neural Networks for E-Commerce.

Area of Science:

• Computer Science
• Artificial Intelligence
• Data Mining

Background:

• Understanding product relationships (complements, substitutes) is crucial for effective e-commerce recommendations and explainability.
• Existing methods often overlook the rich semantic information in the broader topological structure of product graphs, focusing only on directly connected items.
• The connectivity of products beyond immediate neighbors holds valuable data for enhancing relationship prediction.

Purpose of the Study:

• To propose a novel graph neural network framework, the item relationship graph neural network (IRGNN), for discovering multiple complex product relationships simultaneously.
• To leverage multihop relationships within product graphs for improved link prediction and recommendation generation.
• To enhance the explainability of e-commerce recommendation systems through a deeper understanding of product interconnections.

Main Methods:

• Formulating the problem as a multilabel link prediction task.
• Developing the item relationship graph neural network (IRGNN), a graph neural network-based framework.
• Incorporating multihop product relationships by recursively updating node embeddings using neighbor messages.
• Designing an edge relational network to effectively capture relational information between products.

Main Results:

• Demonstrated the effectiveness of the IRGNN framework on real-world product data.
• Validated IRGNN's superior performance, particularly on large and sparse product graphs.
• Showcased the ability of IRGNN to simultaneously discover multiple complex product relationships.

Conclusions:

• The proposed IRGNN framework significantly enhances product relationship discovery in e-commerce.
• Leveraging multihop information in product graphs is key to improving recommendation accuracy and explainability.
• IRGNN offers a powerful solution for complex relationship prediction in large-scale, sparse e-commerce environments.