How Youku Tudou Revamped Its Video Recommendation Engine for Real‑Time Ranking

Background

Youku Tudou hosts billions of videos and daily playback counts, generating massive and diverse behavior data. The recommendation system provides high‑quality video suggestions to users, and the team aims to optimize its effectiveness while reducing the cost of testing new algorithms.

Comparison of Old and New System Architecture

The diagram (shown below) compares the processing steps of the legacy offline‑ranking pipeline with the new online‑ranking pipeline. Although the change appears to be a simple shift of the ranking stage to the online phase, it requires adjustments to data preparation, storage, API encapsulation, and real‑time usage modules.

Advantages

AB testing different models incurs minimal additional development effort and does not require storing a separate set of ranked results.

Ranking data is decoupled from specific models, allowing model selection at request time and enabling lightweight traffic‑strategy adjustments.

Online‑only features enable a one‑pass ranking of candidate videos under real‑time conditions.

Challenges

The serving API must handle a much larger volume of real‑time data requests, such as candidate video sets, detailed feature values, and model weight data.

Online logic must compute ranking results quickly within a very short latency.

High code and service performance requirements are critical to preserving user experience; the architecture team’s effort is essential.

Main Lessons Learned

1. Feature Name Handling Different training dates may produce varying feedback feature sets. The new system performs only offline training; ranking occurs online. To keep ranking independent of specific models, all feature names are globally encoded with unique IDs, and new features receive new IDs.

2. Model Effectiveness Evaluation Offline evaluation is performed first to reduce AB‑test cycles. The process involves preparing one day of data for training and another day for testing, training both control and experimental models, and comparing their global AUC on the test set. If the experimental model’s AUC is higher, it is expected to outperform in online AB tests.

3. Model Weight Fusion Small‑traffic tests showed that merging feature weights from different dates can improve performance. When online ranking uses features computed historically, a newly updated model may lack some feature values for candidate videos, reducing ranking quality. Weight fusion expands the effective feature set and has been shown to increase click‑through rates.

Conclusion

After more than two months of intensive development and testing, the new online‑ranking system was successfully launched. Ongoing optimization will continue, with the team exploring additional methods to deliver an even better viewing experience for users.