Use them train a recommender model; all the docs as processed with SBERT, I use k-means clustering to find 20 clusters and in each cluster it picks out the top 15 scoring articles on “what is the probability i will like this?” using the probability-calibrated SVM on the last 40 days of judgements. I also look at some other lists of articles like: most likely to be voted on HN, most likely to get a lot comments on HN, etc.

I used to run it on a 24 hour cycle and would get around to looking at maybe half of the 300 articles I picked out. Lately I just look at all 300 when I can and then it picks out another 300. So it is showing me better scoring stuff but not so fast to process it.

It ingested maybe 6500 articles this time so it is showing me the “best 5%” while the clustering makes sure I get some articles about sports, some arXiv papers on recommender systems (of course) and other topics.

If you look at what I submit to HN all of it was selected by YOShInOn first and secondly by me so that’s a good sample of what I see except I am into a few things that don’t play on HN (e.g. Premier league football)

Nice work and glad to hear that! If you ever want to chat RSS, please reach out.