Uber calculates passenger ratings using the last 500 rated trips, regardless of when those trips occurred.
In my case, I have 455 ratings, 426 of which are five stars (93.6%). My overall rating is 4.86.
What bothers me isn’t the score itself, but the underlying methodology.
A negative rating from five years ago can carry exactly the same weight as one received yesterday. And if you’re an infrequent user who hasn’t accumulated 500 ratings, that negative feedback could influence your reputation for years.
This raises an interesting question: Should reputation systems measure your historical average, or should they attempt to represent who you are today?
One possible solution is exponential time decay:
Weight = 2^(-t/H)
Where t is the age of the rating and H is a half-life, perhaps 12 months.
Under that approach, a rating from one year ago would carry 50% of its original weight, and one from two years ago just 25%.
Of course, this introduces other challenges. How do you prevent someone with a history of poor behaviour from simply waiting for their reputation to recover? How do you maintain statistical confidence for infrequent users? Would Bayesian smoothing or a minimum effective sample size help?
There’s also a broader philosophical question around the right to be forgotten. Should digital reputation systems indefinitely preserve the consequences of historical assessments, particularly when those assessments are subjective and potentially no longer representative?
I’m curious how others would approach this.
* Is a rolling window of 500 ratings statistically defensible without considering time? * Would exponential decay produce a fairer and more predictive reputation score? * Should platforms expose both a lifetime rating and a recent-behaviour rating? * Are there better approaches already being used in other reputation systems?
I’d be particularly interested in perspectives from people who’ve worked on marketplace ranking, trust and safety, or reputation algorithms.
It seems like a relatively simple mathematical change, but one with potentially significant implications for fairness across digital platforms.