Show HN: HackerNews News Recommender

Will look into it in detail shortly, but would it make more sense to treat the links I upvote as the ones I "liked" rather than the ones I open? Or, is this how it already works?

I don't think so because people oftentimes click on things that may be of interest to them but that they don't want to validate through an upvote.

It's more nuanced than that. I'll often click through a link or comment link and immediately leave. Often times the title was link-bait and I'm not actually interested in the content, or the content was disappointing even if the topic was interesting.

Okiedoke, here's the source: https://github.com/normano/ycChrome

Wasn't there another one of these that worked the same exact way posted a few days ago? http://news.ycombinator.com/item?id=3616763

Similar, but I missed this post.

I thought it was familiar, too ;-)

What do you actually use as features for training? Just the words in the title?

Yup!

You might want to take a look at my implementation. I use the content of the linked sites via viewtext.org. This is probably more reliable.

Doesn't seem to be doing anything for me. Installed it, clicked on a ton of links from hacker news, clicked on "more" a few times, and don't see anything recommended. The extension is definitely installed; I can see the small icon in the omnibox when browsing HN. Is there anything else I need to do?

Try resetting the data from the options. Make sure to close any hacker news tabs while doing so.

Same here. Even after resetting the data, and clicking more again, nothing.

With all of the posts lately, I had a feeling a Chrome extension would be right around the corner.

Does the data not persist between visits? I used this yesterday and it highlighted some posts, but today I got on HN and nothing is getting highlighted.

I'll be releasing an update, some of my decisions were shortsighted, but hey it was a weekend project. :D

Yeah, I'll open source it in the coming days.

How do I know if it's working? I've activated it, clicked a ton of links that interest me and I'm not seeing anything.

In order to train the classifier, you must click the "more" link and make sure the button above in your omnibox is orange.

Click the "more" link how often? Is there a way to tell it that I don't agree with one of its suggestions?

It should only be two times. Two pages of results is enough. And no there is no way you can tell it you don't agree, but that sounds like something I can explore. :)

Do I have to do it every time I need to train it? I never click "more" in regular usage, I read one page and close it.

Yes, you have to. :/ I tried to explore ways to grab the data without needing to click it, but I needed a well defined action (going to the next page is valid to me) that would tell the classifier to train on the data. I'll see if I can find multiple ways to direct certain actions to train the classifier.

Without seeing the source it's hard to ask relevant technical questions, but why not train every time a link is opened?

Online training? I need to read more about it, but I believe there is a partial implementation of it in it weirdly enough.

Dev Source: https://github.com/normano/ycChrome

Naive Bayes is an online classifier anyway, training on click would be the equivalent of clicking a single link and then clicking next. I don't see why it would be hard to do...

Sweet, this is awesome. I'll give you some feedback in a week or two after I give it a spin!

Thanks!

Will you open source it?

You can always unzip the crx file.

Update released!

Doesn't seem to be working for me unfortunately, doesn't show any highlighted links. I think I've trained it properly.

When did you download it, if at the time when I said it was updated try to download it again. What OS and browser version are you using?

Can you go to the extensions page, click the background.html file, inspect elements, console, and type in hnClassifier.getClassObj() and post the output.

Look at the local storage, by typing JSON.parse(localStorage.getItem('classObjs')) and tell me if it looks similiar to the output above.

If you had the older version, please go to the extensions tab and click reset data.

Looks like Linux users had trouble due to case sensitivity. Updated again.

DAYS ago, I remember the other hacker showed his program about HN recommendation written in Ruby.

Same here. When I saw the post, I was wondering if it is the same person, because the other hacker also used the same algorithmic implementation for the recommendation engine.

@apgowz That's the link I was talking about.

Problem there is that I'm far too lazy to label a bunch of data, this one lets you label data with a simple click. :D