With as many YouTube videos that I've watched since 2005, you'd think they might recommend a video with less than 1000 views once in a while. I've found 1 new channel in the last 6 months and used the "Not Interested" option more times than I can count. And the rules are unclear. I don't want tech reviews from 5 years ago, but if I hit Not Interested, does that influence the channel, topic, age, keyword, etc recommendation frequency? I don't mind old DIY or woodworking videos, I'm subscribed to the channel, and watched the recommended video 5 years ago. (Of the current 12 YT Recommended videos, 2 are labelled Watched and only 2 are less than a year old.)
I think part of it stems from the lack of user organization features. Offer too many and you scare away users, while the people with vested interests dump money and time in to wash out any negative opinion (see Amazon reviews). Offer too few and you get poor recommendations during on-boarding/startup.
There are no rules. They just record your preferences and might retrain their black box algo in the future.
The black box decides those rules
Older recommendation systems use background and foreground timeframes to build frames of references, hence repeats.
The question to ask then is: what is the correct thing to optimize for? Dwell time is usually chosen for advertising and stickiness (you are more likely to stick with a service that plays what you like). Optimizing for novelty is difficult because the set of unknown things is much larger than the set of known things. Plus, it is risky from a business point of view.
I am not sure how many actions are needed, but I can think of a few more:
* Not now, but later is fine
* Maybe next month
* Play the opposite of this
* Something similar but new to me
That's not just nice, that is the most obvious "feature" a normal person would design. I've stopped watching Netflix because of nonsense like this. They know I've seen it before, because I watched in their player. Also, even the tiniest, most modest application of analytics would discover that I always watch one series at the time, from start till end, or whenever I'm bored with it. In the past years, I have never ever watched a thing twice. Still, it was constantly showing me things I've already seen.
The only conclusion I can draw from this is that Netflix does not have have my best interest in mind when designing their algorithms and that they don't respect their customers. So I canceled.
With Spotify, they do have a few features that work for me. I've discovered new artists by exploring their "Fans Also Like" feature. The nice thing about this is that they don't try to be too smart there.
Their normal recommendations suffer from the same issues that other sites have and are thrown off by the fact that my tastes are all over the place. I happily listen to sixties psychedelic rock, jazz, metal and some techno or some punk and tend to go from one to the other. Yet I'm very picky about what I listen to. Somewhere along the lines it seems to have decided I'm a middle aged guy (correct) and it consistently does not recommend me any music made this century; which is kind of frustrating if you are trying to find something new to listen to. Recommendation bubbles are a thing and escaping from them is hard.
I work around it by using the fans also like feature and using it's suggested additions to playlists. This works surprisingly well. Example based similarity search is a much simpler problem then recommendations. And it's IMHO a much more interesting feature to explore content with.
Back when I used to use it (about a decade ago), Pandora was good at shuffling, but their library was so small I would get the same songs over and over.
If I can't make it through 5 minutes of a movie I sure as fuck don't want to watch another movie that's just like it. Read the room already.
I don't think that can be stupidity, it must be malice...
I'm dying for human curation.