Hacker News Reader

Comment by irishloop | Hacker News Reader

The problems you're talking about are never going to be solved by ratings companies because you're talking about a completely subjective experience.

There is no such thing as an objectively good restaurant. In my town, there are famous pizza places which people absolutely love and then others who think it's an overrated, pretentious mess.

Nevermind attempting to control for different waiters, chefs, people having a bad day, etc. The problem isn't Yelp-specific.

The problem is entirely human.

View on HN

Foursquare can recommend places based on your personal history of check-ins. Our recommendations for you need not be the same as our recommendations for your friends.

(That said, the venue ratings product we announced today is globally consistent: everyone will see the same rating for a given place. Our Explore product, however, does make personalized recommendations.)

Nevertheless personalized recommendations doesn't avoid the issue that the parent was talking about. Say based on my personal history, Foursquare figures out that I love Pizza. Say I am in Denver, CO tomorrow twenty feet away from two Pizza places, both of which have equal ratings but only one "real": Let us assume that both don't pay Foursquare any money but one is a false positive due to differences in personal taste, artificial pumping of results etc. An algorithm is not going to obviate the issue of making a wrong recommendation.

I think you might be selling it a bit short...

Imagine if they could build up a ratings profile of people who rate the same place similarly, and then network that out... for instance:

Person 1 likes A and B, dislikes C, and hasn't been to D Person 2 dislikes A, likes B, and C, hasn't been to D Person 3 likes B, C, and D, and hasn't been to A Person 4 likes A and D, hasn't been to B or C

So, A has 2 likes, 1 dislike , B has 3 likes, C has 2 like, 1 dislike, and D has 1 like.

That's the start of a rating scale.

But what if an algorithm could identify that, say, Person 1 and Person 4 have similar tastes... so it could recommend D to 1, and B to 4. It can also see that 2 and 3 have similarity, and recommend D to 2.

Now, here's where it gets a bit tricky. The algorithmn can tease out that A and C are opposites - maybe one has great food, but with bad atmosphere/service, and the other is the opposite.

Thus with that deduction, it can recommend B, but not C to 4.

In an ideal world, I would totally agree with you. However, most recommendation systems deal with three issues:

1. Sparsity of data. People surprisingly rate much less than you think they would. In fact, negative ratings are way less sparse than positive ratings (This to me is unintuitive because this is not how I would act but it is what it is).

2. Lack of features for similarity computation. Sometimes, the rating matrix is all you have to compute similarities or you have crappy metadata. You may turn out to be lucky and pull down a facebook open graph and have enough coverage to work with, it depends on your model.

3. The problem of high variance due to latent features (which you alluded to in the last part): Your model gets harder to track due to in sufficient information as to why a place is good or bad. Maybe, there is a correlation between seasonal variations and special cuisines, maybe they had a shitty chef that one time Person 4 came there.

I am not saying it is not do-able, I am just saying it is hard and sometimes ML fairy dust is not enough. :)

Yes, and no. The problems of what movies I might like or what music I like are also entirely human, but Netflix and Pandora have been doing a way better job of figuring that out than Yelp has. And what motivation does Yelp have to do better, anyway-- they don't get their money from me, they get it from venue providers both awesome and terrible.

The motivation is to improve their service or risk losing users like you. They have no business model without happy users.

Netflix and Pandora are largely frictionless experiences. If I spend the money and time to get in a cab in NYC to go to a restaurant that sucks, I'm upset. That's the problem that Yelp is trying to solve. I can't quickly "change the channel" with Yelp, or any sort of ratings for that matter. It isn't a technical or algorithmic problem, it's a human one that doesn't scale well. Quantify != Qualify.

> The problem isn't Yelp-specific. .. The problem is entirely human.

Absolutely. So one part of the solution is to stop putting restaurants on an "objective" 1-5 scale, and averaging every human together.

Instead, cluster restaurants so you can "people who liked the overrated, pretentious mess also liked X..."

Or using a backend weighting system for computing total score for a location.

Low weights: new users, numerous reviews (spamming), low rated reviews

High weights: older users, high rated reviews

Yelp actually does have an algorithm for computing the final star rating. It is not a simple average across all ratings. It takes contributor status on Yelp, age of review and other signals into account before the final score is published. You can, of course, dig into the rating distribution to see the spread between 1 and 5 but I doubt that most users go that far.

Google tried this, but to a first approximation nobody used/uses Hotpot/Google Places/Google+ Local.

That's actually pretty easy: "People like you rated this place X".

Yay machine learning! Though to scale that is a nontrivial issue...

A not-quite-as-good but easier statement to make: People who like place X also like places Y, Z... etc.

Also, sparsity of data. The fact that people are precious unique snowflakes. :)

http://en.wikipedia.org/wiki/De_gustibus_non_est_disputandum

In my town, there are famous pizza places which people absolutely love and then others who think it's an overrated, pretentious mess

Let me guess, you live in Phoenix?

Reply on news.ycombinator.com