4 ms·

That's an interesting project. I've been exploring this from a different angle, in the form of generated email newsletters (https://www.findlectures.com/form?ty

That's an interesting project. I've been exploring this from a different angle, in the form of generated email newsletters (https://www.findlectures.com/form?type=alert https://www.findlectures.com/form?type=alert). This uses the contents of the articles as well - I'm crawling links on programming subreddits. If you start from keywords you can use NLP (Word2vec) to measure how close articles are to your interests, so for instance "python, machine learning" gives you stuff on tensorflow, scikit learn, etc whereas "java, machine learning" gives you articles on spark, Deeplearning4j. You can also measure how similar articles are to each other - getting the most dis-similar articles avoids the "piling on" problem mentions.

pmoriarty 9y ago
Many years ago, back when Bayesian spam filtering was the hot new thing, I wrote up a little program that would classify articles in RSS feeds based on whether I found similar articles to be interesting or not interesting in the past. The ultimate goal was to save me time and effort in manually classifying them, as I and everyone else do when we scan through what we come across on a daily basis. Instead of manually doing that, the hope was the program could do it for me, and I could just focus on reading the interesting articles and not even have to deal with the uninteresting ones. From that experiment I learned a few things: - First, that I'd have to manually scan through all the articles anyway, just in case the classifier made a mistake and maybe dumped an article I found intensely interesting in the uninteresting pile. - Second, that having to consciously think about which article was interesting or uninteresting in order to do the training, about whether the classifier was working or not and which articles needed to be reclassified, about having to re-train it when it messed up, and so on was a hell of a lot more work than just scanning through my RSS feed manually and deciding on which articles I found interesting or not myself. - Third, my interests were not static things that the algorithm could learn and classify on correctly from then on out. My interests were constantly changing. Sure, maybe there were a handful of things I always found interesting or uninteresting -- but overall what I found interesting or not changed from day to day. It was also kind of unpredictable, even to myself. The third point kind of argues towards the approach of the HN front page, which is un-classified and un-tagged. I've read all sorts of great articles on HN that if I'd been going by some pre-written list of interests that I had, I would have never have read. I do still often wish for tagging on HN anyway, just because there are certain types of articles that I really never ever want to read, and I'd love to be able to exclude them. But the vast majority of HN articles aren't of that kind (or I wouldn't be here). That experiment with bayesian classification turned out to be rather short-lived, as I found the whole thing way too much of a bother to maintain and to retrain when it misclassified articles. I'm still reading RSS feeds the old fasioned way today, and am a little suspicious about any AI/machine-learning-like approaches to article classification.
- garysieling 9y ago
  Yeah, that makes a lot of sense. I started this from making an email list where I'm sending "lunch and learn" talks which I manually curate / write up descriptions. About half the people like that style, and half have specifically requested something more topical. There are a bunch of great curated email newsletters around specific interests (Javascript Weekly, etc) so I'm aiming to do something similar, but more granular. So far the ML thing has been promising, but it helps to start from a pre-vetted dataset.
- crznp 9y ago
  > First, that I'd have to manually scan through all the articles anyway, just in case the classifier made a mistake and maybe dumped an article I found intensely interesting in the uninteresting pile. But I'm sure there are interesting articles that didn't get classified at all. It doesn't seem like the end of the world if you lose a little wheat with the chaff. > overall what I found interesting or not changed from day to day This seems like the problem. If you rated an article highly yesterday, it doesn't mean you want to read the same article again today. "Interesting" largely means "novel" and it is hard to find that by looking at similarities to what was new in the past.
  - pmoriarty 9y ago
    "It doesn't seem like the end of the world if you lose a little wheat with the chaff." But is it losing a little or a lot? No way to tell without looking. Even if it misses just a little, that little bit might have been crucial. If it misclassifies "NYC NUKED!!!" as uninteresting that's a single mistake that could make me oblivious to a hugely consequential event. Of course, as a human classifier, I'll doubtlessly making my own mistakes, and maybe the AI classifier could help me out by pre-classifying articles for me, but it could also be misleading and will cost me in terms of spending time on training and maintenance. I'm not really sure what the right solution is here.
    - PaulHoule 9y ago
      If NYC got nuked, I'd probably hear about it some other way.

pmoriarty 9y ago

Many years ago, back when Bayesian spam filtering was the hot new thing, I wrote up a little program that would classify articles in RSS feeds based on whether I found similar articles to be interesting or not interesting in the past. The ultimate goal was to save me time and effort in manually classifying them, as I and everyone else do when we scan through what we come across on a daily basis. Instead of manually doing that, the hope was the program could do it for me, and I could just focus on reading the interesting articles and not even have to deal with the uninteresting ones. From that experiment I learned a few things: - First, that I'd have to manually scan through all the articles anyway, just in case the classifier made a mistake and maybe dumped an article I found intensely interesting in the uninteresting pile. - Second, that having to consciously think about which article was interesting or uninteresting in order to do the training, about whether the classifier was working or not and which articles needed to be reclassified, about having to re-train it when it messed up, and so on was a hell of a lot more work than just scanning through my RSS feed manually and deciding on which articles I found interesting or not myself. - Third, my interests were not static things that the algorithm could learn and classify on correctly from then on out. My interests were constantly changing. Sure, maybe there were a handful of things I always found interesting or uninteresting -- but overall what I found interesting or not changed from day to day. It was also kind of unpredictable, even to myself. The third point kind of argues towards the approach of the HN front page, which is un-classified and un-tagged. I've read all sorts of great articles on HN that if I'd been going by some pre-written list of interests that I had, I would have never have read. I do still often wish for tagging on HN anyway, just because there are certain types of articles that I really never ever want to read, and I'd love to be able to exclude them. But the vast majority of HN articles aren't of that kind (or I wouldn't be here). That experiment with bayesian classification turned out to be rather short-lived, as I found the whole thing way too much of a bother to maintain and to retrain when it misclassified articles. I'm still reading RSS feeds the old fasioned way today, and am a little suspicious about any AI/machine-learning-like approaches to article classification.

garysieling 9y ago
Yeah, that makes a lot of sense. I started this from making an email list where I'm sending "lunch and learn" talks which I manually curate / write up descriptions. About half the people like that style, and half have specifically requested something more topical. There are a bunch of great curated email newsletters around specific interests (Javascript Weekly, etc) so I'm aiming to do something similar, but more granular. So far the ML thing has been promising, but it helps to start from a pre-vetted dataset.
crznp 9y ago
> First, that I'd have to manually scan through all the articles anyway, just in case the classifier made a mistake and maybe dumped an article I found intensely interesting in the uninteresting pile. But I'm sure there are interesting articles that didn't get classified at all. It doesn't seem like the end of the world if you lose a little wheat with the chaff. > overall what I found interesting or not changed from day to day This seems like the problem. If you rated an article highly yesterday, it doesn't mean you want to read the same article again today. "Interesting" largely means "novel" and it is hard to find that by looking at similarities to what was new in the past.
- pmoriarty 9y ago
  "It doesn't seem like the end of the world if you lose a little wheat with the chaff." But is it losing a little or a lot? No way to tell without looking. Even if it misses just a little, that little bit might have been crucial. If it misclassifies "NYC NUKED!!!" as uninteresting that's a single mistake that could make me oblivious to a hugely consequential event. Of course, as a human classifier, I'll doubtlessly making my own mistakes, and maybe the AI classifier could help me out by pre-classifying articles for me, but it could also be misleading and will cost me in terms of spending time on training and maintenance. I'm not really sure what the right solution is here.
  - PaulHoule 9y ago
    If NYC got nuked, I'd probably hear about it some other way.