Deep Learning Practices for Personalized Recommendation at Meitu: From Recall to Ranking

Meitu's community recommendation system serves over 100 million monthly active users, delivering personalized content through a waterfall‑style feed and subsequent similar‑item feeds, with the goal of understanding both content (visual, textual, behavioral) and user profiles (demographics, device, history) to drive multi‑objective optimization.

The production environment faces three major challenges: (1) massive scale with millions of candidate items per day, (2) strict real‑time latency requirements (<300 ms) for billions of ranking requests, and (3) long‑tail distributions of users and items that demand robust generalization.

In the recall stage, Meitu employs three deep‑learning‑based strategies: Item2vec learns item embeddings from short‑term co‑occurrence using a skip‑gram with negative sampling model; YouTubeNet incorporates user embeddings to retrieve personalized candidates via FAISS; and a dual‑tower DNN jointly models user and item features, enabling real‑time user‑side vector updates.

For ranking, the system evolved from a linear‑regression baseline to Neural Factorization Machines (NFM), then to Field‑wise Neural Factorization Machines (NFwFM) that add field‑wise bi‑interaction, matrix‑factorization and FM modules, reducing model size while boosting CTR by ~5 %. A multi‑task NFwFM further shares a hard‑parameter backbone for click‑through and follow‑conversion tasks, using homoscedastic uncertainty weighting to prioritize click prediction and achieve a 15.65 % lift in conversion and 1.57 % CTR gain.

Overall, the integrated deep‑learning stack—covering both recall and ranking—delivers consistent improvements in click‑through rate (up to 4.72 %), user dwell time (up to 2.98 %), and conversion metrics, demonstrating the effectiveness of embedding‑driven retrieval and advanced multi‑task ranking models in a production‑scale recommendation system.