4 ms·
I totally agree. 4.6 as a cutoff is insane. I regularly rate things 3 if it's good, 4 if great (on Google Maps). What Uber should have is: Rate your ride: 5
by jbarberu 7y ago
I totally agree. 4.6 as a cutoff is insane.
I regularly rate things 3 if it's good, 4 if great (on Google Maps).
What Uber should have is:
Rate your ride:
5 - great
4 - bad
3 - really bad
2 - awful
1 - shouldn't be allowed to drive ever again
- colejohnson66 7y agoWhat I think they should do is just have one questions with two answers: “Would you ride with this driver again? Yes or no?”
- mywittyname 7y agoAll 5 star scales operate as follows: 5 - Good 4 - Terrible 3 - Terrible 2 - Terrible 1 - Terrible
- jbarberu 7y agoThat seems to have become the norm. To me that's just people misunderstanding how to use a gradual scale. I recently got a phone call from my apartment management after I left a bunch of 3s and 4s on their annual review, they seemed distraught they got such bad scores. From my point of view they were doing just fine. The fact that they're not excellent in every regard shouldn't be surprising when I'm not paying through my nose to live there, I'm perfectly happy with 3.5/5 since that's what I'm paying for.
- kerng 7y agoIt's an ordinal scale with arbitrary meaning. 4 doesn't mean its twice as good as a 2 star rating (there is no interval or ratio). It basically means 5 = all okay! 4 and below there were issues of varying subjective degrees.
- ghaff 7y agoSee also: https://xkcd.com/1098/ https://xkcd.com/1098/ Ratings tend to be skewed in various ways everywhere. But it's particularly extreme when people know the norm is to give a perfect score for any transaction that didn't have serious flaws--especially when you know lower scores have material negative real-world implications for individuals trying to earn a living.
- BrandonM 7y agoThat basically is what they have (see bottom): https://www.uber.com/en-AU/blog/how-ratings-work/ https://www.uber.com/en-AU/blog/how-ratings-work/ I think it's absurd that they try to dictate user behavior based on broad averages, as opposed to normalizing each individual's ratings to fit a standard distribution, then averaging those normalized ratings.