Confounder Balance in Next Basket Prediction.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Next basket prediction (NBP) is crucial for e-commerce, aiming to forecast user purchases based on historical data.
• Current NBP methods often overemphasize frequently interacted items, leading to biased predictions.
• Item interaction frequencies represent average preferences but fail to capture individual user interest levels.

Purpose of the Study:

• To address the bias in NBP caused by prioritizing high-interaction items.
• To develop a method that accurately models individual user interests for improved NBP.
• To propose a novel model that mitigates bias while preserving personalized user preferences.

Main Methods:

• Utilized causal analysis to demonstrate the bias in existing NBP approaches.
• Introduced user-specific weights based on repeated interactions to model user interest.
• Developed the confounder balance prediction (CBP) model using counterfactual inference.
• Employed counterfactual inference to isolate and balance confounders (item frequency and user interest).

Main Results:

• The proposed CBP model effectively mitigates bias in NBP.
• CBP preserves and refines individual user interest modeling.
• Experiments on four real-world datasets show CBP outperforms state-of-the-art methods.
• CBP achieves significant advantages in prediction accuracy and relevance.

Conclusions:

• Existing NBP methods are theoretically biased due to over-reliance on item interaction frequencies.
• User-specific interest modeling and bias mitigation are essential for accurate NBP.
• The CBP model offers a robust and effective solution for unbiased and personalized next basket prediction.