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.