Personalized Product Recommendations Using Convolutional Neural Networks to Drive Sales

Personalized recommendations are a major component of modern e-commerce experiences. For emerging product categories with rapidly evolving catalogs, recommendation quality often depends not only on prior purchases of similar users but also on the customer’s existing product ecosystem and contextual purchase patterns. In addition, exposure driven by large-scale promotional campaigns can introduce noise into interaction data, making recommendation modeling more challenging. To address these issues, we developed a convolutional neural network recommender that captures n-gram-like groupings of products in customer purchase histories to model ecosystem structure more effectively. We also incorporated L∞L_∞-norm regularization to improve robustness against noise introduced by high-volume promotional campaigns. Our analysis shows that the proposed model improves both prediction accuracy and recommendation diversity compared with baseline approaches. We further report results from multiple online experiments across retail placements, where the production system consistently achieved statistically significant lifts in upsell metrics. Finally, we examine how external factors, including catalog changes and seasonal shopping periods, affect the performance of the deployed system.