3 ms·

I still prefer a Gaussian Mixture for this unless you're saying you roughly even N per cluster

PaulHoule 2y ago
It all depends on what you want. If 3/k of your items are about soccer or 2/k are about cryptocurrency then you get 3 and 2 clusters for them respectively whereas on some level you might want 1 cluster for each of them and also want one cluster for a topic which with prevalence 1/3k. On the other hand, having 2 clusters for a topic with 2/k could be seen as fair on another level... one way or the other I appreciate the system not having too many parameters and not being too sensitive to those parameters and I'd expect it to do roughly the right thing if k is set a little too high. Early one I had a fight with the system where it kept showing me articles about soccer despite hating soccer because I had roughly 1/k articles about soccer (an RSS feed from The Guardian!). I wound up thinking a lot about feature engineering for sports and started reading soccer articles in detail and pretty soon I was amazed by games that were 0-1 and an own goal or that went 8-0 and pretty soon I became one of those people who is watching soccer at 9am on Saturday. Related to that is the question of “should it show me more articles from clusters that I like better?” which defeats the point of diversity but certainly would improve most quality metrics (though I’d get sick of arXiv papers about recommender systems, which was most of what I got when k=1 in the beginning, and would be screaming for articles about cricket or something instead.) I haven’t done a lot of work to try to improve that system because I like the feed it makes already and I’m not sure what to optimize to make it better. If it has a weakness it is that it has a lot of latency but there are certain things like sports articles and Ken Shiriff’s blog which are more timely and I’d like a way for some articles to “jump the line” without seeing too many of those articles overall…. But I’m not sure how to quantify what success is.

It all depends on what you want. If 3/k of your items are about soccer or 2/k are about cryptocurrency then you get 3 and 2 clusters for them respectively whereas on some level you might want 1 cluster for each of them and also want one cluster for a topic which with prevalence 1/3k. On the other hand, having 2 clusters for a topic with 2/k could be seen as fair on another level... one way or the other I appreciate the system not having too many parameters and not being too sensitive to those parameters and I'd expect it to do roughly the right thing if k is set a little too high. Early one I had a fight with the system where it kept showing me articles about soccer despite hating soccer because I had roughly 1/k articles about soccer (an RSS feed from The Guardian!). I wound up thinking a lot about feature engineering for sports and started reading soccer articles in detail and pretty soon I was amazed by games that were 0-1 and an own goal or that went 8-0 and pretty soon I became one of those people who is watching soccer at 9am on Saturday. Related to that is the question of “should it show me more articles from clusters that I like better?” which defeats the point of diversity but certainly would improve most quality metrics (though I’d get sick of arXiv papers about recommender systems, which was most of what I got when k=1 in the beginning, and would be screaming for articles about cricket or something instead.) I haven’t done a lot of work to try to improve that system because I like the feed it makes already and I’m not sure what to optimize to make it better. If it has a weakness it is that it has a lot of latency but there are certain things like sports articles and Ken Shiriff’s blog which are more timely and I’d like a way for some articles to “jump the line” without seeing too many of those articles overall…. But I’m not sure how to quantify what success is.