Hacker News Reader

Comment by jmduke | Hacker News Reader

Ideally -- and take this with a huge caveat, as I'm not exactly in the aggregation business -- measure it by clicks, comments made, and time spent on page/bounce.

Overall, I'd like to see aggregators move to the Pandora model, rather than the Hype Machine one. With Pandora, I can specify a jumping off point (genre, artist, song) but it still refines my tastes from there.

Hype Machine, I've 'liked' hundreds of songs but I still need to manually peruse through specific blogs to find new songs I like. It's unideal.

View on HN

I believe that in the future that's the way to do it, but not right now.

Besides, I'm not a fan of automatic recommendations as they have a lot of drawbacks. Their major problem is: they won't help you find something you didn't know you liked.

Their major problem is: they won't help you find something you didn't know you liked.

This is patently untrue. It is more likely to find you stuff that you didn't know you like than the lowest common denominator strategy.

Are you talking about any existing recommendation engine that I can try out to prove myself wrong?

Just how exactly will they do it? I'm really interested, maybe I'm missing something?

Simply speaking, if a system knows all I listen to is classical music, it can recommend me a new classical composition, but it cannot all of a sudden recommend me dubstep (which I might actually like).

We're talking about Digg doing this as a differentiation. It is not commonly used because it is very heavy in IO and CPU for individual users (though a site like reddit could take a shortcut just by periodically ramping up some spot high CPU instances and batch generation correlations between users on every subreddit, then staling that until the next run), and sites like Reddit and previously Digg ran on stacks that greatly limited their ability to scale for this.

Regarding music taste, if the system knows that you have a strong correlation with a number of users who like classical but periodically explore new music, what they like will likely be something you might be interested in. This, in practice, is exactly how music discovery happens in the real world (that the discoveries of the people you share musical tastes with are more likely to interest you).

What you are talking about is still incapable of recommending (with a high chance of the recommendation being spot-on) new things. In music a new thing would be a completely different genre, not a new song in your favorite genre.

The system you are describing will find people who have a strong correlation with me and who also like:

1) Classical music and Justin Bieber.

2) Classical music and Metallica.

3) Classical music and Dubstep.

And it will recommend me Bieber, Metallica and Dubstep. I might like Dubstep, but I won't like Bieber and Metallica. How is it better than listening to random songs?

Reply on news.ycombinator.com