Musinsa Boosts Personalized Push CTR by 21.3% with Unified Embeddings

The Musinsa Core-E Personalization team built Layer 0 (Unified Embeddings), a user representation layer shared by all models and decision-making processes, and applied it to personalized push notifications, achieving a 21.3% increase in CTR.

Previously, the company operated dedicated models for each service such as recommendations, push notifications, and coupons, leading to structural inefficiencies and learning silos due to repeated training on the same users. To address this, they adopted an approach that integrates user, product, brand, and category learning within a single common space.

Knowledge Graph-Based Unified Embedding Design

Musinsa utilized a Knowledge Graph and the TransE model to reflect the connected structure of the data. By representing entities and relations as vectors, various actions such as clicks, likes, and purchases are learned in an integrated manner within a single space.

• Training Scale: 128-dimensional embedding space, 12 types of entities, 23 types of relations, and approximately 150 million edges trained
• Data Cleansing: Excluded users and products with no meaningful responses to increase the information density of the graph

Application Effects and Validation

Unified embeddings function as a common asset rather than dedicated models for individual services. When applied to the category preference model, CVR improved by up to 7.8%, confirming that performance enhanced even when only the user representation method was changed while maintaining the existing model structure.