Maybe it's kind of there in how instances interact, or the post visibility features, or some of the forks like Hometown, but Reader had that right-feeling combo of simplicity, personalization, and built-in network.
Maybe it's kind of there in how instances interact, or the post visibility features, or some of the forks like Hometown, but Reader had that right-feeling combo of simplicity, personalization, and built-in network.
The technology to do this existed circa 2000 but I’d never seen anyone give it a serious try. One reason for that I realized is that, for a centralized service, you can get better recommendations based on other people’s preferences.
Assuming you have a huge user base, users will perceive a collaborative filtering interface learns better. My current classifier takes 20 or 50 judgements to start to understand that I like news about the NFL but I don’t want to see articles about that other kind of football.
The content-based model produces a good relevance score (“do I care about the topic?”) but it has no indication of quality apart from that. (Consider how Google is ranking both on query-document match and a quality score.). One input to my system is computer science papers on arXiv and it shows me more papers on ML than I can possibly read, never mind seriously comprehend the abstract in all cases. The relevance score does a great job of vanquishing theoretical CS papers I can’t stand but I wish I saw the 30 best ML papers and not the 30 most tropey. It is gratifying when it shows me papers about how to build this kind of classifier!
After all this I have a lot more respect for collaborative filtering despite how dangerous it is socially.