I appreciate the categories, even if it wasn't personalized.
One idea for future development might be to add a point score based on Naive Bayes (which has been discussed previously). Not as a filter per say, but just as an indicator.
Sure, you can use a bookmarklet or write your own thing, but I've always thought it'd be an interesting idea to have a centralized like/hate server, where you could add different sources, say HN. That might be a different project, but adding a score based on the likelihood of "I'd read this" (not recommendation as I should note) would be a interesting in terms of skimming.
Let's say one of your users is in a hurry, he might want to skim the most (homogenous) likely news that'd interest him/her, whereas if it's late at night in leisure time, one might be more interested in new/abnormal stories.
Wow, that's an awesome feedback and great link! Thanks for your kind words! That's a pity there's no API access to user's saved stories (as far as I know). This would be great source for Bayes training. Maybe a chrome extension could train Bayes on HN visited links.
I agree that it's a difficult issue, as some might feel it's an intrusion of privacy if that data is used. Other might feel it's annoying to have to up/downvote stories both on HN and on your site. That's a trade-off. Perhabs both, or allow import of saved data, so users can make a dump from HN and import it in your site.
I think the other HN api used to have that (if people trust them to pass their user/pw to HN safely), but seems to be disabled: http://api.ihackernews.com/.
So yeah, the easiest way to start it would be to have people up/down vote on your site, or use a heuristic measure such as: if you click a link, you "like" it, if you're shown a link, but don't click a link you "ignore" it (which should be less negative than hate, as there can be a number of reasons for not clicking a link).
As in:
I get a list of 10 items from each category, of those 10, which do I click, which don't I. So once an item is shown to me - save it as "bad" by default in the DB.
If I click it, change it to a good result.
You'd might have to change the weight more than this, but it should be a good start, and requires the least amount of effort for the user.
I played with machine learning and HN stories a while ago. All my classifiers only managed to predict stories I would click on with around 60% accuracy.
When I tried to classify stories I would click on myself on two different occasions I also only managed around 60% accuracy.
So I figure what I feel like reading depends on a lot of internal context too.
In the end I got good enough results feeding stories with more than X points into my instapaper automatically. Few false hits and I never know about the near misses!
I'm doing this now as a side-project. I need quite a lot of data, unfortunately, and Scikit-Learn is giving me some trouble, but hopefully I'll have something I can show you guys soon.
Sure, as soon as I have something that works! I'll probably give it another stab over the weekend and see if it can predict any better than chance.
If you want to do something useful on this for yourself, save the post IDs you liked and ones you didn't like. I don't currently have an interface for that (I use the command line to vote), but I can import them very easily from a list of IDs into your account later on.
I like it, but I wouldn't mind an option to customize the line spacing and font colors (alternating colors, maybe?) - it's pretty dense to skim through.
After installing the extension, just open Stylebot and copy the following design for skimhn.com: http://pastebin.com/5Nupsm3B. it's still rough, but it's a start :)
That way you don't have to operate with logins and everyone's setup is share-able. (Then you can look in the future of letting the creator define the url.)
One idea for future development might be to add a point score based on Naive Bayes (which has been discussed previously). Not as a filter per say, but just as an indicator.
E.g. http://joelgrus.com/2012/02/16/hacking-hacker-news/
Sure, you can use a bookmarklet or write your own thing, but I've always thought it'd be an interesting idea to have a centralized like/hate server, where you could add different sources, say HN. That might be a different project, but adding a score based on the likelihood of "I'd read this" (not recommendation as I should note) would be a interesting in terms of skimming.
Let's say one of your users is in a hurry, he might want to skim the most (homogenous) likely news that'd interest him/her, whereas if it's late at night in leisure time, one might be more interested in new/abnormal stories.
Just a though. Good job.