Personalized Restaurant Ranking with a Two-Tower Embedding Variant

A lightweight variation of a two-tower embedding model was developed to improve personalized restaurant ranking in a food delivery app, addressing the failure of simple popularity-based sorting. The model uses a simplified two-tower structure with a frozen language model for restaurant embeddings and represents users based on their contextually filtered interaction history. To handle sparse data and capture current user intent, the model averages embeddings from past orders matching the current search tag and uses multi-task learning to predict clicks, adds-to-basket, and orders. This approach resulted in a significant conversion rate uplift and proved generalizable to other recommendation surfaces within the app without requiring retraining for new item collections.