Meitu's Personalized Recommendation System: Architecture, Features, and Optimization Strategies

In the era of information explosion, users face difficulty discovering relevant content while long‑tail items receive little exposure. To address this, Meitu has built a personalized recommendation system that connects its massive user base with a huge collection of images and videos, significantly improving user experience.

Meitu's recommendation technology was presented at the Meitu Internet Technology Salon, where expert Cai Qisen shared the team's practice and exploration in personalization.

Overview

Personalized recommendation aims to link users with content, enhance experience, and optimize the content ecosystem. The system is deployed primarily in the Meipai app across three scenarios:

Live streaming – real‑time online ranking.

Meipai Hot (homepage) – a Feed‑style endless scroll where users explore content.

"You May Also Like" on video detail pages – recommending similar videos.

The core principles are to understand both content and users.

Understanding Content

Content understanding is essentially feature extraction for short videos, divided into four categories:

Basic features – duration, resolution, tags, etc.

Visual features – gender, age, height of people; scene and object recognition.

Text features – titles, captions, comments; mined for keywords, topics, sentiment.

Interaction features – likes, comments, shares, plays, aggregated into statistical metrics.

Understanding Users

User profiling includes natural attributes (gender, age, city), social attributes (occupation, marital status), and especially interest attributes, which may extend beyond the platform (e.g., gaming, shopping).

Recommendation Process

The pipeline consists of three stages:

Recall stage : Reduce billions of videos to a few hundred candidates using statistical recall, simple models, graph calculations, and deep‑learning‑based recall.

Estimation stage : Apply machine‑learning models that jointly model user‑context‑item dimensions (e.g., LR, GBDT, NFM, DCN) to predict relevance scores.

Ranking stage : Combine novelty, diversity, and accuracy to produce the final ordered list presented to users.

The overall architecture is divided into OFFLINE (data collection, cleaning, labeling, feature engineering, model training), NEARLINE (real‑time stream processing, feature updates, incremental model training), and ONLINE (serving engine that performs recall, fusion, and ranking).

Timeliness

Because user interests shift rapidly, Meitu emphasizes recommendation timeliness:

Recall timeliness : Maintain a real‑time similar‑video (I2I) index; update it based on user actions to generate instant recall results.

Estimation timeliness : Adopt online learning (e.g., online gradient descent) for models such as LR, while addressing instability with techniques like FTRL (Follow‑the‑Regularized‑Leader). The FTRL update rule and pseudo‑code are illustrated in the original article.

Online learning is integrated with a Parameter Server architecture, where FeatureServer ingests Kafka streams for real‑time feature computation, and TrainServer generates training samples for model updates. A Balance Cache controls the consumption rate of positive/negative samples to keep the data distribution stable, achieving second‑level model and feature updates.

Cold‑Start Solutions

Both user and content cold‑start are tackled:

User cold‑start : Multi‑level feature combination recall using any available user information (demographics, context) and subsequent estimation with enriched features (gender, age, occupation, time, location, video tags) to predict click‑through, watch time, and retention.

Content cold‑start : Treat as an Exploration‑Exploitation problem. Use multi‑armed bandit strategies (UCB, Thompson Sampling) to allocate traffic among random Explore, personalized Explore, and personalized Exploit. New items receive visual classification and MAB scoring before being placed into exploratory candidate pools.

Content Diversity

To balance accuracy, novelty, and diversity, Meitu re‑models the multi‑objective problem, enforcing diversity and novelty constraints while maximizing accuracy. Content similarity is defined via tags, visual classification, or behavior‑based clustering.

Content Clustering

User behavior sequences are used to cluster content similarly to topic modeling. LDA treats each user as a document and viewed videos as words, producing topic distributions that serve as content clusters. Alternatives include Word2Vec embeddings or DSSM‑based vectors; the DSSM network diagram is shown in the source.

Future Outlook

Meitu plans to explore more scenarios, provide richer personalized services, shift from supervised learning to reinforcement learning for expected reward estimation, and further refine the content ecosystem.

The article concludes with a call to register for the 10th Meitu Internet Technology Salon, focusing on classic video architecture and optimization cases.