Weibo Recommendation Algorithm Practice and Machine Learning Platform Evolution

Weibo, a leading Chinese social platform with 240 million daily active users, continuously upgrades its recommendation algorithms to extract valuable information from massive data and enhance user experience.

Recommendation Scenario : The system recommends related content after a user reads a post, aiming to personalize content consumption and guide user interest convergence.

Machine Learning Workflow : Data from item side (posts) and user side (behaviors) are processed, stored in material and data middle‑platforms, and used to generate training samples for model training, which then influences recall and ranking stages.

Samples and Features : Sample strategies include noise reduction, benchmark labeling, and filtering inactive users. Features are categorized into ID, reader, author, context, environment, and material features, with ID features showing the most significant impact.

Algorithm Practice : Recall uses FM and dual‑tower models with user and item embeddings, while ranking evolves from simple rules to LR, FM, FFM, Deep&Wide, DeepFM, and FiBiNET, achieving notable AUC improvements.

Scaling Challenges : Model size leads to storage, computation, and gradient explosion issues. Solutions involve a custom parameter server (weips), sparse computation, low‑precision parameters, gradient clipping, and batch‑norm techniques.

Deployment with Weiflow : Models are deployed via a configurable workflow that automates resource handling, enabling rapid deployment and a 35 BP online metric gain.

Weibo Machine Learning Platform : The platform has evolved to support trillion‑scale parameters, million QPS, and minute‑level model updates. It offers one‑stop services for data storage, computation, sample management, feature services, training, and model serving.

Platform Services : Compute service unifies heterogeneous clusters; sample service manages data and permissions; feature service provides feature catalogs and security; training service integrates with model libraries; model service enables one‑click online deployment and version control.

Impact : The platform’s integration has driven significant KPI improvements for recommendation and mobile products, demonstrating the value of large‑scale machine learning in production.

Weibo aims to leverage its technology to make the world better, and invites the community to engage through sharing, likes, and follows.