New User Product Recommendation for Q-commerce via Hierarchical Cross-Domain Learning
Bhavi Chawla, Shubha Shedthikere, Jairaj Sathyanarayana · 2024
In this work, we address the new user cold start problem in the product recommendation system for a quick-commerce (q-commerce) grocery delivery service. Traditional recommendation systems built on user-product interaction data tend to perform poorly for new users owing to data sparsity. This has led to the emergence of cross-domain recommendation systems, which leverage the data from a relatively richer ‘source’ domain to improve recommendations in the target domain. In our case, online food delivery, being one of the more evolved services of our platform, is the source domain to the online grocery delivery referenced above. This enables us to leverage the data from the food domain. There is a large body of literature on cross-domain recommendation systems which typically involves learning a cross-domain mapping function between the customer or the item embeddings in the two domains and leveraging that mapping function to derive the embeddings for new users in the target domain. We show that such approaches are sub-optimal where sales distribution is long-tailed, making the embeddings noisy. This is further aggravated when applied to q-commerce settings where location-specific geographical and cultural diversities have to be considered. Given these nuances, we propose a neural network-based hierarchical cross-domain mapper, which leverages the inherent hierarchy in the item taxonomy and learns multiple mapping functions between customers’ food category preferences and product category preferences in the grocery domain. These category-level preferences are then used to personalize product recommendations in the grocery domain through a learning-to-rank model. We show that the proposed algorithm outperforms the embedding mapping-based approach and the popularity-based baseline by 30% and 8% respectively, in terms of NDCG in an offline experiment and 4% improvement in conversion in an online A/B experiment which is significant at our scale.