[PaulHoule]

I’ve been experimenting with a smart RSS reader that is, I think, successful with news articles. It trains a text classifier to filter content I like. I added the RSS feed for HN comments to it and thought it was too much, but I still have it ingest the ‘top comments’ and like some of what I find, I think the highest probability score it’s assigned to a comment is around 0.5 though, contrasted to 0.9 that I regularly see for some machine learning papers and articles about the Ukraine war.

This is not really the right approach to comments because they are embedded in a tree, the parent and children should matter for classification. I think though a comment viewer would be possible and I also think metafilter would be a good source of comments, possibly better than the metafilter articles.

(Oddly in the last month I have been thinking about inferred preference scores as more ‘real’ in some sense than revealed preferences. HN articles that have a high comment/score ratio are frequently knockdown-dragout matches and by training a classifier on that ratio I have a ‘clickbait’ score that expects a lot of comments on ‘Is Bitcoin a Shitcoin?’ but very few on anything that has to do with graph databases.)