Kurly Achieves Up to 140% ARPU Increase with Collaborative Embedding Recommendation System Using Knowledge Distillation

Kurly's Data Service Development Team built a personalized recommendation system based on Collaborative Embedding that processes user behavior like language. To overcome the limitations of existing search- or popular-item-focused recommendations, they adopted an approach that interprets users' 'silent actions' such as clicks and add-to-cart events using Transformer context understanding technology.

Semantic ID and Dual-Head Architecture

Product text information is compressed via RQ-VAE to generate Semantic ID, which is represented by separating it into Content Space and Collaborative Space. The model applies a Dual-Head Architecture to balance accuracy and speed. The Generative Head generates the Semantic ID of the next ordered product and understands complex context, while the Retrieval Head handles high-speed search (ANN) based on vector similarity.

Performance Optimization through Knowledge Distillation

When initial experiments showed the Retrieval Head's performance was lower than the Generative Head's, Knowledge Distillation was introduced. Soft Label information such as purchase probability from the Generative Head (Teacher) was transplanted to the Retrieval Head (Student), resolving the Representation Bottleneck problem. This enabled the model to reflect subtle preference differences beyond binary judgments.

Online A/B Test Results

Clear results were confirmed in an A/B test conducted on 50% of total traffic. On the 'Sale' page, add-to-cart (ATC) increased by approximately 32%, and revenue per user (ARPU) rose by approximately 140%. The 'Discount' page also recorded approximately 12% ATC increase and 32% ARPU increase, proving that improved personalization precision directly translates into revenue.