[mg]

I often think that marketplace rating should work like this:

Every 3 rides, you get asked "Which one of the last 3 rides did you like the best?".

Then the driver you select gets 1 star. So an average driver would get 0.3 stars per ride. For display purposes, this can be normalized to whatever scale is most comfortable for users.

This should completely get rid of the problems with the popular "rate every seller right after the transaction" approach.

[Majromax]

> Every 3 rides, you get asked "Which one of the last 3 rides did you like the best?".

Note that this is a variation of stack ranking, which has incentives for the usual sorts of manipulation like avoiding "difficult" times like rush hour.

[mg]

One could say that a "rush hour ride" is a different service than a ride at a less crowded time.

Hotels in a small town might have better reviews than hotels in New York. But if you want to stay in NY, you compare between the hotels there. So the hotels in the small town do not have a benefit of their "manipulation" to offer rooms in a less crowded city.

[gabereiser]

But in this case, they are comparing the same trip at a different time and traffic congestion level. Make this the “difficulty multiplier”.

[rjmunro]

Or just "was this ride experience better or worse than your previous ride" for every ride apart from the first one you ever take. You could then use the Elo rating system to track driver scores.

[mg]

When every transaction is rated, the buyer will still get nagged by the seller to give them the star.

My thought behind giving one star to one of the last 3 rides is that the seller does not know when the buyer will give out that star. It might be 3 months in the future. So it seems futile to nag the buyer about what they will do then.

Calculating an ELO value should also be possible for more than two players.

[WithinReason]

Then you just say better, every time. It's the same problem as the 5 star system.

[rjmunro]

If anyone does this, their ratings will just average out to nothing. You could detect it and filter these people out.

You could reduce bias towards the most recent trip by adding a delay before asking.

[wink]

What if the only time I've ridden an Uber was in 2014? That's an actual example but even with 3 taxi/uber rides per year I probably would not remember.

[joking]

that's in line with the past week elo everything ( https://news.ycombinator.com/item?id=36393406 ) submission. There are many statistical ways to improve the rating systems, but if you don't act and proactively filter bad actors, is not a zero sum game.