Use ML to improve Shorts discovery, metrics, safety & cold start

As the Product Manager for YouTube Shorts, the core strategy is to use Machine Learning to create a powerful, self-improving ecosystem. The primary goal is to enhance content discovery through a sophisticated deep learning recommendation engine. This model would move beyond basic signals to develop a true understanding of content and user preference. It would ingest multi-modal signals from the videos themselves—analyzing visuals, audio, and text—and combine them with nuanced user interactions like dwell time and, critically, how quickly a user skips a video. By optimizing for long-term user satisfaction instead of just immediate engagement, we can deliver a more diverse and compelling feed that keeps viewers coming back.

To ensure this engine is effective, we would employ a dual-layered approach to metrics. The first layer consists of our North Star product metrics, such as user retention, session duration, and creator success. The second layer involves the technical ML metrics, like offline precision and recall, which are validated through rigorous online A/B testing. A new model is only successful if it demonstrably improves the end-user experience. Parallel to this, we would use ML to guarantee platform safety at scale. This involves proactive, multi-modal classifiers that scan content at the time of upload for policy violations, as well as NLP models to filter toxic comments and systems to intelligently prioritize user reports for human review.

Finally, we would address the classic cold start problem with a two-pronged ML strategy. For new users, the system would use an “explore-exploit” model, learning their tastes rapidly from their very first interactions. For new videos from creators, which have no interaction history, the system would rely on content-based analysis to identify the topic and style. It would then serve the video to a small, highly-relevant seed audience. If this audience responds positively, the main recommendation engine then amplifies the video’s distribution, ensuring every creator has a fair chance to find their audience and grow.

• This topic was modified 1 month, 4 weeks ago by Ayushmaan Dwivedi.