Personalized Push Notification System: Embedding, Recall, and Ranking Techniques at Meitu

The article introduces push notifications as a crucial user‑engagement tool and explains how deep learning advances enable more effective, less intrusive personalized push strategies.

Embedding Evolution : It covers three embedding approaches—Word2Vec‑based feed embeddings, Airbnb‑style listing embeddings that incorporate user interactions beyond clicks, and recent graph neural network embeddings that combine user behavior with item features.

Recall Models : Four recall categories are described: global recall (hot lists, keywords), personalized recall (user behavior‑driven), attribute‑based recall (demographics, device), and content‑based recall (similarity, keyword, and copy similarity). Specific models such as YoutubeNet, clustering recall, and various copy‑recall methods are detailed, with reported click‑through improvements.

Ranking Models : The ranking pipeline progresses from a simple logistic regression model to increasingly complex architectures: xNFM (enhanced factorization machine), dual‑tower, triple‑tower, and finally a field‑wise three‑tower model that integrates recall‑source information, achieving cumulative click‑through gains up to 23.9% and an additional 5.68% uplift.

Conclusion : Embedding serves as the foundation for recall and ranking, while the layered recall and advanced ranking models together form a robust, scalable push recommendation system that significantly boosts user engagement.