Next Basket Recommendation with Basket Size Prediction

A next basket recommendation model that incorporates basket size prediction, named NBR-WBS (Next Basket Recommendation with Basket Size).

Description

• The embedding module includes a basket encoder using attention mechanisms for item importance and a basket-size encoder using Long Short-Term Memory (LSTM) for basket size embeddings.
• The basket preference module combines basket and size embeddings, processed by a shared Transformer for multi-task learning to generate the final state vector.
• The prediction module utilizes the final state vector of baskets for two tasks: predicting basket items and sizes using separate predictors. The results from these predictors are then combined to provide users with dynamically sized basket recommendations.

Datasets

• Ta Feng
• Dunnhumby
• Instacart

Contributions

• To the best of our knowledge, this study is the first to take into account basket size dynamically in the next basket recommendation.
• The proposed NBR-WBS model serves as a framework for multi-task learning. It not only forecasts the probability of items within the baskets but also predicts the size of users' baskets at different time points.
• Extensive experiments were conducted on three real datasets, demonstrating that the proposed NBR-WBS model outperforms existing methods.