Inside Baidu Mobile’s Personalization: Recommendation Engine & Cloud Architecture

Theme 1: Personalization in Mobile Baidu

Mobile Baidu is a super‑app that goes beyond search to cover many aspects of users' lives. It faces challenges such as rising information‑acquisition costs, diverse user needs, and the need for scenario‑based, personalized operations. The goal is to make each user’s Baidu experience unique, enabling fast and effective discovery of the most relevant information and services.

Personalization relies on rich user tags (gender, age, occupation, interests, etc.) and purpose tags (intent, focus, demand). These tags support card recommendations, cold‑start mitigation, and targeted operational activities. For example, post‑95 users show higher interest in games and anime, while pre‑95 users prefer shopping and food.

The recommendation system starts from traditional collaborative filtering (user‑based and item‑based similarity) and combines item similarity with content features. Baidu’s implementation uses an item + content hybrid similarity measure.

The current recommendation architecture is loosely coupled via APIs, built on Hadoop for machine‑learning and data‑mining, ingesting statistical and behavioral data and outputting recommendations through a recommendation API.

Key challenges include massive user scale (tens of millions of offline model trainings per week) and integrating diverse content sources, while advantages stem from Baidu’s abundant ML resources and extensive search and profiling data.

Mobile‑specific demands—limited screen space, higher precision, richer contextual signals (time, location, scenario)—drive the need for more accurate recommendations compared to traditional PC‑based systems.

Evolution of the system moves from rating‑based to ranking, page optimization, and context‑aware recommendations, incorporating algorithms such as point‑wise and pairwise learning‑to‑rank.

In the O2O domain, Baidu integrates third‑party services (e.g., Nuomi, maps) and applies expert‑crafted matching rules as prior probabilities, later refined by actual click data, improving performance by about 0.5 % and further boosting results by 2 % when scene features are added to machine‑learning models.

Personalization extends beyond recommendation; Baidu explores personalized cards on the home screen, requiring continuous discovery of new user features, intents, and life‑stage signals, and matching them with product attributes.

Theme 2: Cloud‑Side Architecture Design and Practice

Mobile Baidu’s cloud side faces high concurrency, large capacity, flexibility, reliability, and compatibility challenges due to its hybrid browser/CS model. The solution is a multi‑layer architecture: a resource‑aggregation layer (cloud resources, caches, multiple DBs), a middle layer (dynamic UI, account state, multi‑protocol support, proactive delivery), and an upper layer (configuration, operations, H5/NA coexistence, dynamic client updates).

Traditional LAMP stacks are adapted for the super‑app, focusing on performance optimization, compatibility, extensibility, and cloud controllability.

Performance optimization addresses high QPS, peak concurrency, and unpredictable writes, using techniques such as page caching, cache negotiation, and request merging.

Compatibility and extensibility are achieved through passive client‑side feature detection, code branching, resource acquisition ordering, and specialized logic to handle multiple client versions and platforms, improving cache hit rates and efficiency.

Cloud controllability enables dynamic content, events, and configuration delivery without requiring client upgrades, allowing the server to adjust UI, behavior, and policies in real time.

The overall cloud architecture resolves long software lifecycles, uneven traffic, unstable resource links, and complex requirements.

Theme 3: Front‑End Exploration at Mobile Baidu

The front‑end team uses SASS, Zepto, OpenJs, and an EMD module system, providing solutions for templating, local caching, and SPA development.

Engineering is driven by a unified FIS integration solution with custom plugins for dependency analysis, code linting, on‑demand loading, compression, and merging, producing resource maps that guide runtime module loading.

Publishing leverages a Tiny platform for configuration management, modular APIs, and monitoring via the MDP device‑profile platform.

Future front‑end directions include hybrid development models, pre‑installing code into the app, and exploring React Native approaches.