Returnformer: A Graph Transformer-Based Model for Predicting Product Returns in E-Commerce.

Area of Science:

• E-commerce analytics
• Machine learning
• Graph theory

Background:

• High product return rates increase costs for e-commerce retailers.
• Lenient return policies and impulsive purchasing contribute to return rates.
• Accurate prediction of returns before payment is needed for proactive management.

Purpose of the Study:

• To develop a novel model for predicting e-commerce product return behavior before payment.
• To reduce potential financial losses for retailers through proactive return management.

Main Methods:

• Proposed Returnformer, a novel return prediction model based on the Graph Transformer.
• Integrated global topological embeddings to mitigate structural information loss.
• Employed Graph Transformer for long-range user-item dependency capture within subgraphs.
• Introduced a graph-level attention mechanism for global return pattern propagation.

Main Results:

• Returnformer demonstrated superior performance compared to four machine learning models.
• Achieved higher prediction accuracy on a real-world e-commerce dataset.
• Outperformed existing state-of-the-art models in return prediction.

Conclusions:

• The proposed Returnformer model enables accurate prediction of product return risks before payment.
• Retailers can implement timely and proactive preventive interventions to reduce losses.
• The model supports proactive return management in e-commerce.