1
$\begingroup$

I notice that Netflix, which I think used to use a five-star scale for rating content and give predicted ratings for unrated content on the same scale, now just has basic like/dislike buttons. Music streaming services like Pandora and Spotify seem to employ the same approach. My question is what makes a binary response superior to a likert-type scale when it comes to generating the best recommendations?

Is it related to the nature of the algorithms (e.g., classifications are more accurate when there are fewer possible responses)? Or is the advantage more to do with human psychology (easier for users to accurately choose between like/dislike as opposed to love/like/neutral/dislike/hate)?

$\endgroup$
  • $\begingroup$ I suspect the answer is yes: it is both easier to learn a binary classifier, and the binary response captures enough of the variance in the signal. You would have to delve into the data to get a definitive answer. $\endgroup$ – Emre Jun 19 '18 at 0:04
1
$\begingroup$

This article from Business Insider answers your question from the business perspective.

They mention that the original problem is that the way Netflix used the 5-star rating was not standard within the industry:

Netflix’s Cameron Johnson, who oversaw the shift, told Business Insider that it all came from the realization that Netflix had always used star ratings differently than the rest of the internet, but that this distinction wasn't clear to users.

And the way ratings were presented to the user discouraged them to contribute. Their solution was the thumbs up/thumbs down approach because it was more clear to the user that they were training an algorithm:

So when looking for a replacement, Netflix wanted to make sure that was clear. That’s why Netflix settled on “thumbs up/down," which is widely understood to imply that you are training an algorithm to know what you like, Johnson said.

I suspect there is more than that thought, and maybe binary data is simply "good enough".

This chapter from Mining of Massive Datasets has some interesting considerations on ratings versus boolean utility matrices.

From page 339:

If the utility matrix is not Boolean, e.g., ratings 1–5, then we can weight the vectors representing the profiles of items by the utility value. It makes sense to normalize the utilities by subtracting the average value for a user.

Specially when treating rating matrices one need to be aware that different users may have different understanding of ratings, hence you will most likely normalise the matrix before calculating the recommendations, while for binary data this is not required.

| improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.