Insights into iQIYI's Recommendation Platform Architecture and Practices

As business growth creates many scenarios requiring recommendation services, teams often lack domain expertise and sufficient manpower to build complete recommendation systems. iQIYI addressed this by launching a recommendation middle‑platform, achieving over 30% improvement in development efficiency and completing the build and launch within just 10 days. Senior Technical Manager Zhang Shijun shared the full technical layering of this platform.

Understanding middle platforms: A business middle platform provides generic business capabilities (e.g., e‑commerce, recommendation, search), while a data middle platform consolidates scattered data into unified user profiles. A technical middle platform offers common technical services such as big‑data processing, storage, and middleware, though its classification as a “platform” can be debated because it is not directly tied to specific business functions.

Middle platform vs. platform: A platform is a broader concept that can include both front‑end and back‑end. A middle platform is a specialized form of platform, focusing on a subset of services. For example, an e‑commerce platform encompasses the entire stack, whereas an e‑commerce middle platform only provides the core business capabilities. Underlying technical services (big‑data processing, storage, etc.) are also considered platforms that support many middle platforms.

iQIYI’s timeline and motivation: Planning for a recommendation middle platform began in the second half of 2018. Prior to this, the recommendation team’s limited resources forced many business units to build their own ad‑hoc recommendation systems, leading to high labor costs and sub‑optimal results. The middle platform was created to enable any front‑end business to connect easily and quickly to a unified recommendation service.

Organizational adjustments: A dedicated recommendation middle‑platform team was formed, and internal presentations were held to promote the platform to other teams. Because the platform must serve diverse business needs, higher requirements were placed on overall architecture design and big‑picture thinking.

Technical layering of the recommendation middle platform: The platform relies on foundational services such as a content production platform, content data warehouse, user‑behavior log platform, data processing platform, storage platform, machine‑learning platform, and analytics platform. Internally it consists of: - Recommendation content pool - User‑behavior collection - Feature engineering - User portrait generation - Recall algorithm repository - Ranking algorithm repository - Online inference engine - Reporting dashboards - Configuration platform Each module is described briefly, emphasizing its role in abstracting business complexity and enabling configurability.

Case study – efficiency gains: Before the middle platform, supporting a new business required weeks of scheduling and full‑time staff. After the platform’s launch, a new image‑text feed product integrated the recommendation service in just 10 days, with each module taking 1–2 days and minimal human effort, resulting in a performance boost of over 30%.

Challenges and lessons learned: The main difficulty was designing for universality. Different businesses have varied recommendation entities, behavior‑log formats, and requirements. The solution involved a unified data warehouse, abstracted entity fields, configurable mapping tables, and a standardized behavior‑log schema that can be automatically transformed or manually adapted.

Future evolution: iQIYI plans to further automate the integration process to eliminate manual configuration steps and to continuously improve recommendation effectiveness by incorporating internal optimization results and the latest algorithms, ensuring all business units benefit from enhanced recommendations.

Guest introduction: Zhang Shijun is a Senior Technical Manager at iQIYI, specializing in recommendation engine architecture, high‑concurrency distributed systems, and the ongoing construction of iQIYI’s recommendation middle platform.