Causal Inference–Based Recommendation Algorithms for User Growth in Video Platforms

In the era of mobile internet, user growth has become the top priority for applications, shifting from the traditional AARRR framework to a RARRA approach that emphasizes retention, re‑activation, and sharing; for content‑driven products, personalized algorithms are crucial to drive these outcomes.

The article reviews three successful growth models in the video domain: the "Headline & Kuaishou" model (e.g., Toutiao, Douyin, Kuaishou) that builds a complete content production‑consumption ecosystem; the "Qutoutiao" model that applies game‑like milestone incentives and virtual/real rewards to acquire mass users; and the "iQIYI/Tencent Video" model that relies on heavy content acquisition and large‑scale funding, resulting in limited algorithmic impact.

For long‑video platforms such as Youku, user growth must be tackled across content supply, distribution, rights design, and product design. From an algorithmic perspective, the goals are twofold: (1) deep user‑state modeling to identify intervention factors that move users from low‑ to high‑value states, and (2) upgrading personalized distribution by applying these models across multiple scenarios to meet and stimulate video consumption needs.

Alibaba Entertainment has built a user‑growth system comprising push notifications, DSP external traffic acquisition, and new‑user acquisition, as well as a revenue‑growth system based on marketing and advertising. Central to both systems are a causal‑inference‑based recommendation algorithm and an uplift‑model‑based marketing gain model, which have already delivered notable performance improvements.

Traditional data analysis focuses on correlation rather than causation, limiting the design of effective interventions. Causal inference, by estimating true causal effects and counterfactuals, enables the construction of unbiased user‑cf models that match low‑activity users with high‑activity counterparts using unbiased features such as demographics, installed long‑tail apps, and active search queries.

The proposed unbiased user‑cf design differs from classic user‑cf in two key ways: it excludes passive recommendation data (push, in‑app, and operational recommendations are not used), and it only matches low‑activity users to higher‑activity users, avoiding matches among users with similar activity levels.

Business results show that after deployment, the algorithm achieved over 50% lift in both CTR and click volume for silent users in push scenarios, and nearly doubled re‑activation volume in DSP campaigns, demonstrating the practical value of causal inference in user growth.

Looking forward, the current applications address only two user‑state transitions; however, the methodology can be extended to the full spectrum of user states illustrated in the membership conversion diagram, promising broader impact across data analysis, product design, and distribution optimization for the industry.