3 ms·
The wrong solutions the article describes are certainly wrong. It doesn't really make much of a case that the particular solution it proposes is particularly go
by gjm11 17y ago
The wrong solutions the article describes are certainly wrong. It doesn't really make much of a case that the particular solution it proposes is particularly good.
For instance, here's another much simpler formula that avoids the problems described in the article: score = (pos+1)/(pos+neg+2). It's the posterior mean of Pr(random person likes the product), if your prior is uniform on [0,1]. You can adjust this to favour items with more ratings by, e.g., using a prior that "prefers" smaller values of that probability; one easy way to do that is to pretend that every item starts with a certain number of positive reviews and a certain number of negative reviews; you end up with a formula of the form (pos+A)/(pos+neg+B).
Is the article's formula better than this? Probably, but the author hasn't said why.
- tel 17y agoIf we're going to use a Bayesian treatment, the real sin is throwing away your uncertainty. What I would love to see is a ranking algorithm that provides an elegant, intuitive interface to fuzzy ranking. If we can't be confident of the ordering, why pretend? It's a usability problem, not a statistics problem. Your estimator is slightly more resilient than Amazon's and somewhat less conservative than the author's. It's definitely simpler, but who cares? Visitors never have to compute it after all.
- araneae 17y agoAgreed. Even if you find the formula scary, he implemented it in Ruby for you. It's a pretty short trip to implement it in the language of your choice.
- gjm11 17y agoI think it's a huge usability problem, if we're talking about something Amazon-like that's presenting ranked data to an unsophisticated audience. Why pretend? Because it produces something the user might be able to understand, that's why. (Yes, it would be extremely cool to make an elegant, intuitive interface to a fuzzy ranking. One reason why it would be extremely cool is that it's totally not obvious how it can be done, or even that it can be done.) I wasn't suggesting that my estimator is better because it's simpler, nor that anyone should care; just pointing out that the author of the article made a huge leap from "here are two things that don't work" to "and here is the specific arbitrary math-heavy thing I think you should do instead".
- tel 17y agoI agree. I think it's a great problem to motivate a startup on.
- ntoshev 17y agoWith Bayesian average (http://news.ycombinator.com/item?id=1219154 http://news.ycombinator.com/item?id=1219154 , http://wiki.answers.com/Q/What_does_true_Bayesian_estimate_mean_in_connection_with_the_IMDb_Top_250_ratings http://wiki.answers.com/Q/What_does_true_Bayesian_estimate_m...) you can easily put a slider saying how much the missing votes count (or how much preference do you give to items with more ratings).
- Predictor 17y agoThis solution is called a "Laplacian correction", and I agree with your recommendation.