I'm curious, what are the main factors that determine algorithms? Is user interaction dominant or is content analysis? For example, when I don't like some of the recommended videos, how does the algorithm respond? In your opinion, how personalized is this system?
How does YouTube's algorithm determine recommendations?
👁️ 10 views💬 4 replies❤️ 0 likes
4 Replies
Ever since I started paying close attention to how YouTube’s recommendations work, I’ve been really struck by how videos keep popping up even if you don’t like them or disable them somehow. What’s most convincing is that it doesn’t just track "likes" or clicks—it also looks at how long you watch each video. For example, I noticed this with some movie analysis channels: if I only watched 10 seconds and skipped it, soon after the recommendations shifted to cooking or travel content.
That said, when you hit "Not interested" or "Don’t recommend this channel," you *do* see a change—but the algorithm takes a little while to adjust. Still, random content that doesn’t fit you keeps slipping through, especially if it’s something you’ve searched for before. In the end, even though it’s personalized, it still has a pretty big margin of error.
YouTube's algorithm is actually quite similar to the way our scattered thoughts work. After my mind opened up like a bag of sugar, I tried watching a video on a random topic and realized how the system captures that. Last year, I searched for "オーディオ機器レビュー" (audio equipment reviews) in an old Japanese electronics magazine. Within a few weeks, ads related to that started appearing on YouTube. Once, my homepage even recommended "JVC KD-X380BT sound system repair guide"—which I never clicked on. Strangely, the algorithm didn’t stop suggesting new videos on that topic and kept reminding me about it over time. It felt like the system was saying, *"You didn’t like it, but you’re still interested."*
I also noticed how my dislike for certain videos affects the algorithm. For example, when I close an ad after just two seconds, I eventually stop seeing that same ad on different channels—but I also start encountering fewer free videos on that topic. So the system doesn’t interpret it as *"the user hates this"* but rather *"maybe they lost interest."* Similarly, after watching a Minecraft mod review but ignoring the following GTA modding videos that popped up, the related recommendations decreased over a week—but didn’t disappear completely. The algorithm doesn’t judge based on a single action; it compares multiple interactions and continuously updates a "personal profile."
Honestly, I was almost shocked by how personalized the algorithm is. One morning, I checked YouTube on my friend’s phone and saw a completely different world in their feed. While my recommendations were full of audio systems and game reviews, theirs had dozens of J-pop and manga suggestions. It’s like the system builds a unique database for each user, updating it in real time. When I think about how much data must be behind such a detailed and dynamic system, my mind goes completely blank.
I've always been curious about this topic, especially since I've run some experiments to test how "smart" the algorithm really is. From what I've seen, two major factors come into play: user interaction (likes, watch time, comments) and content-based recommendations (keywords, topic similarity). For example, when the algorithm suggests videos that go beyond what I usually watch, I can't help but think about how well it actually understands the content.
I've noticed something in my own experience: for a while, I was watching fitness videos, and then suddenly the algorithm started recommending "furniture assembly" videos. At first, I just brushed it off with a "whatever," but after watching and liking a few, the algorithm quickly corrected itself. The situation you mentioned about "videos I didn’t like at all" is also interesting—you can really see how personalized the system is based on user reactions. A friend of mine had a similar experience; after skipping 3-4 videos in the same way, the algorithm shifted to a completely different topic. So, it seems like the system is really pushing the user in a certain direction.
I think the algorithm mostly considers watch time and click-through rate—if I close a video right away, it starts reducing recommendations from that category after a while. I do sometimes get suggestions that don’t interest me, but generally, when I close the ones I don’t like, the system seems to take note and adjust.