How does YouTube's recommendation system build its model based on the videos we watch? What weight does it give to factors like watch time and likes? Also, how do the algorithm's preferences change when directing users to a new channel or different type of content? In your opinion, how effective is user feedback within this system? What are your thoughts on the transparency and control of the algorithm?
How does YouTube's recommendation algorithm work, and why does it sometimes suggest unexpected videos?
👁️ 77 views💬 2 replies❤️ 0 likes
2 Replies
Last week, while looking for a Vue.js tutorial, the algorithm first showed me several cooking videos because I had spent more time on a recipe playlist. But as soon as I clicked on a coding video and watched over 50% of it, the suggestions shifted to dev tutorials and even emerging channels. It really showed me how much watch time matters more than likes, and how direct feedback (clicks, pauses) quickly reshapes the system.
YouTube's recommendation model fundamentally operates on a "watch history + engagement" cycle. The signals combined—such as watch time (CTR and completion rate), interactions like likes/comments, and metadata like video titles/tags—create a "watch profile." From my experience, if you finish a video without leaving it (especially past 70%), the recommendation weight for that topic spikes immediately; if you half-watch and hit "skip," the system treats it as a weak signal and reduces similar video suggestions.
When switching to a new channel or content type, the "exploration" phase kicks in. If you introduce variety in a short span (e.g., 5–10 minutes of unrelated topics), the algorithm interprets it as a "shock" signal and pushes new recommendations—but doing this too often creates "signal noise," degrading recommendation quality. A practical tip: if you want to test a new theme, watch 3–4 videos in that niche at 80%+ completion and like them, then return to your usual content with a "buffer" (2–3 mid-tier videos) in between. This way, the algorithm registers the new theme as "prioritized but consistent," reducing unexpected videos in your recommendations. User feedback (likes/dislikes, "Not interested") is also highly effective—even a couple of dislikes can reduce a category’s visibility, giving you active control over your recommendations.