Avoid Strategic Mistakes—How to Properly Set Optimization Goals for Recommendation Systems

This article discusses the problem of setting reasonable optimization goals for recommendation systems, emphasizing that algorithm engineers often focus on model innovation while neglecting the strategic question of whether the chosen objective truly serves business needs.

As a well‑known internet figure said, “Don’t let tactical diligence hide strategic laziness,” a principle that applies equally to technology adoption: a technically advanced model is useless if its optimization target diverges from actual business requirements.

For most commercial companies, the recommendation system’s goal should be derived from the company’s commercial objective. The article uses YouTube and Alibaba as case studies to illustrate this principle.

YouTube’s use of “user watch time” as an optimization target is explained: because YouTube’s revenue comes from video ads, which are proportional to how long users watch, the platform optimizes for watch time rather than click‑through or play‑rate metrics. Optimizing for watch time encourages recommending longer, higher‑quality videos, whereas click‑rate optimization would favor short, sensational titles.

Figure 1: YouTube recommendation system output layer.

YouTube’s ranking model treats recommendation as a classification problem that predicts whether a user will click a video, then extends it to predict watch time.

In contrast, e‑commerce platforms like Alibaba aim for purchase conversion. The user journey is abstracted into two steps: (1) product exposure and click, (2) purchase on the product detail page. Because the commercial goal is purchase, the appropriate model is a CVR (conversion‑rate) estimator.

The article points out a mismatch between the training space (click + conversion data) and the prediction space (exposure stage), which can cause biased estimates. A naïve solution is to build separate CTR and CVR models, but this does not optimize the ultimate conversion goal.

Alibaba’s solution is the Entire Space Multi‑task Model (ESSM), a multi‑objective model that simultaneously simulates exposure‑to‑click and click‑to‑conversion stages. The lowest embedding layer is shared between CTR and CVR to alleviate CVR’s data sparsity, while separate towers learn pCTR and pCVR; their product yields pCTCVR.

Figure 3: Alibaba’s ESSM architecture.

Thus, both YouTube and Alibaba design their recommendation‑system objectives to reflect true business goals and to align training and serving scenarios, a principle that should guide any recommendation‑system design.

The article also stresses that defining optimization goals is an interface problem with other teams (product, operations, content). Clear, shared objectives allow technical teams to focus on model improvement while product teams can align features with commercial strategy, reducing friction and strategic errors.

Finally, readers are invited to join the “DataFunTalk Recommendation Algorithm” community for peer discussion; to join, add the WeChat contact “DataFunTalker” and reply with “推荐算法”.