A good recommendation system for music needs to take into account proximity in play order to other songs you've ranked when you rank it, and their tempo and genre, time of day, etc.
Ever wonder why Netflix keeps on pushing bad content ?
> "This past week, Netflix officially debuted their new Amy Schumer stand-up comedy hour The Leather Special. And it was instantaneously met with negative reviews. Some claim that Schumer's biggest critics got on Netflix and purposely drove down the rating of the special, some without even watching it. Schumer herself blames the 'Alt-Right' for sabotaging her latest effort. Now, it is being announced that Netflix is ditching its five-star rating system for something much more streamlined, a rating system that owes itself to the legacy of the late Gene Siskel and Roger Ebert. Did Amy Schumer have anything to do with this new rating system? Probably not. But once in place, it will certainly help her Leather Special find a more appropriate audience."
https://movieweb.com/netflix-cancels-5-star-rating-system/
'Probably not', but it sure seemed like coincidental timing to me.
Business of course wants to keep churn low, and they think time spent on the site is the best way to do that.
That (and "people mostly don't use ratings beside 1 and 5") is at least the reasoning I've read every time I've seen this topic mentioned in blogs/talks.
I do also feel Netflix is pushing it's in-house content way more these last few years, to the detriment of their recommendations though.
The second issue is a lot of people use it like 1/5 system. While yes some users use more of the intermediates and you can adjust on per user basis for that feature it’s just not as useful as you’d expect.
Of course any decision like this I recommend a company ab test and I’m just summarizing past experience in this area.
I had a similar discussion at a prior job on having creators add content tags to there submission. The main avoidance was while good tags would be nice for recommendation, if tags are required less content would be submitted and that hurts the platform.
That is great! This means the people giving feedback would tend to be power users that are able to giver higher quality feedback.
> As a side effect short content is very nice as finish/playtime become strong signals that require a user to do almost nothing.
Perfect! It is clear by now that users vastly prefer short content that gets to the point quickly.
So basically 5 stars ranking would vastly increase the quality of the content but that is bad, how?
The only issue I see is (political) vote brigading. Maybe set additional hurdles like minimum account age and the like for voting. Strong content moderation could help. I guess there is no silver bullet for that issue.
Even for power users it may not be net benefit for them. Recommendations are not based solely on that user’s ratings but all ratings. If power users give better signal but total quantity of signals decreases then it’s likely that recommendation quality drops for both casual/power users. Making up guess numbers a 5 star system is likely to get you 3x less data points and many of those data points are of similar quality anyway (users that only rate 1/5 stars). Worse signal that has higher volume often leads to better model.
I later suggested ways to make the hurdles for voting higher, which would fix the problems you mentioned. Required account age, maybe even how much content you need to have watched. Honestly, we could just filter out the people that only do 1/5 star votes. (With niche content of course there would be the issue of having no/just a few votes but voting would not be the only metric anyway.
I am very heretical towards common held UX believes and think trying to design for the lowest common denominator just results in all around awful experiences for everyone. We should rather strive to empower users.
That is quite unlikely. What you call power user is someone distinguished by willingness to rank. It does not make that person better representation of all users or be more knowledgeable off quality or more experienced. Just more judgy.
It means you will be recommended based on tastes of special minority and likely alienate the rest of users in the process.
As someone responsible for designing and incorporating end-user activity/feedback into ML, I have to tell you this is exceedingly unlikely.
The number of power users, especially on a large consumer app like Spotify, is dwarfed by the broad userbase (I estimate ~1% of Spotify users at most qualify as "power users").
Would I rather have power user-only feedback, or 100x the feedback? Definitely definitely 100x the feedback from all users.
I can also say working at companies like that generally any system changes require an ab test and long running back test. I’ve generally seen companies move from 5 star to simpler like/dislike system and those changes would have been ab tested and found that user experience is overall better. I don’t know for specific systems whether overall better was across all users or some users.
-1 -> dislike, +1 -> like, +2 -> favourite, +4 -> heart
Labels are arbitrary. How might this weighting affect learning?
Slice open an orange along its equator. Looking at this slice you have a shape kind of like an *
Each of these wedges can have a different tag. Those wedges facing up would have a positive value, and down a negative.
The site the orange slice voting system is used on can easily decide the number of wedges that make sense and the tags associated.
Instead we get binary or unitary choices that don't really mean much.
Edit: Image for example
+2 love
+1 like
0 don’t mind it
-1 dislike
-2 never again
I'd have to think an intelligent recommendation engine could do a ton more with star rankings than a simple up/down too. 5=love, 4=like, 3=ok, 2=dislike, 1=hate. It could even be optional; they could just treat like and dislike as 4 and 2 or something, and let people who want more precision use them.