Next Basket Prediction with Deep Learning

Challenge

Data-driven product recommendations are an excellent strategy for retail companies to activate customer-specific revenue and margin potential. Recommender systems analyze historical baskets to uncover patterns and similarities between customers or baskets. Our client, a national retail company, faced the challenge of combining customer and basket-specific information in a recommender system to utilize both information sources for generating recommendations.

Approach

First, historical transaction data was processed to extract product and basket information for each customer. These data were then processed in a recurrent deep learning model. The advantage of such models is their ability to natively consider customer, product, and basket-specific factors simultaneously. Additionally, various benchmarks, such as naive approaches like Last Basket or simple machine learning models, were calculated to better evaluate the performance of the new approach.

Result

The new deep learning-based approach demonstrated a significant improvement over the naive models. At the product level, the prediction of the next basket was four to five times more accurate compared to naive or simple model-based suggestions. The recommendation system is now used in various online and offline marketing applications.