Data-based Recommendations

bytepawn.com

9 points·Maro··8 comments

Most sites do recommendations as a big Venn diagram. If my_upvoted_items intersection your_upvoted_items > x then recommend me your_upvoted_items union my_upvoted_items(complement).

I don't think this is a great approach. If you and I have similar tastes, chances are that I have intentionally avoided the things that you like but that I have not voted on.

I'm working on a music recommendation site that, I hope, will avoid this problem.

That would be a nice improvement for Hacker News.

I assume that it is possible to track my votes and compare them with other users. If I voted for an article, I should get a recommendation for another article that people with similar votes have found interesting.

This would also give an additional incentive to use the vote button, since it would improve my experience.

I agree. Similarly, if IMDb would have good recommendations, I'd cast far more votes just to get better recommendations.

Isn't this what the Netflix prize is all about? Based on a history of user rentals and ratings you want to find recommendations.

Unfortunately, Netflix is not available where I live, so I can't comment on it.

data >> algorithms

Which is why I asked: http://news.ycombinator.com/item?id=265728

My point in this regard is not so much about data vs. algorithms. It's the fact that even the sites that use the data they have suck.

I think that the much, much larger user base of Amazon that is voting with dollars rather than diggs is the big issue.

Reply on news.ycombinator.com