Really like the vision approach to classifying web pages, I've been thinking google should add this to their algo for a while (if they havent already).
Classifying individual parts of pages (as Diffbot seems to be doing) is difficult, but I suspect google could take screenshots of pages reported as spam or whatever as one class and compare those to screenshots of pages w/high pr to get a pretty interesting classifier they could use as an extra datapoint. Could be an interesting experiment anyway, using data they've got lying around.
Wanted to thank the HN community for all your encouragement. I first released the Diffbot API as a "Show HN:" post last year (http://news.ycombinator.com/item?id=2310852). $2M+ and lots of hard work later, we're powering some of the largest destination sites out there like Stumbleupon and the new Digg.
Congrats! Diffbot offers very compelling technology. Incidentally, the infographic you created was gorgeous. Do you mind sharing who you used to create it?
Glad you liked that! That infographic was made by the multi-talented Dan Mayer, who we highly recommend. Check out his portfolio at http://danmayer.com
The visual analysis method looks very interesting Mike! Will you identify or extract information from pages with product info, events, food recipes or product reviews?
wow, pretty cool. I wonder though is there much use for it outside of aggregate sites like digg? Even for a site like reddit, all the content is already split up into categories by users. While this is really cool, I'm not really seeing much use for it. What are some problems that this will solve?
The most pressing problem that the page classifier solves is for the current community of developers using Diffbot. We only offer APIs at the moment for extracting information out of frontpages and article pages (see http://www.diffbot.com/our-apis/). While we'll be releasing more soon for the other page types that you see in the infographic chart, this limitation is a problem if, for example, you're passing in a photo page or recipe into the article API--you're not going to get an article back since its not appropriate. Having the technology to visually understand what the types of pages are allow us to route those requests.
Beyond that, it's useful for companies that perform analytics or use the information to do better search or auto-categorization.
Are you going to be releasing an events API in the near future? I've had an app idea banging around for a few months that this API would be perfect for.
The next API we'll train an API for is likely image/photo pages. Event pages are definitely on our roadmap, but we've still got work to do there. Could definitely use your help with training data and learning more about your use case.
How does caching works?
Is there any focus on security?
Multiple geolocations?
I liked the TOS :)
----
Diffbot.com is made available for personal, non-commercial, and commercial purposes. Services are provided as-is, and we do not make any guarantees on the quality or performance.
Conceptually I like the product, it's something I would consider paying for. But in practice it doesn't seem to perform that well. It misclassify things it should get right (article hosted on posterous; a youtube page; hacker news) and for some queries it just returns results for a completely different webpage.
Thanks. Yes, the page classifier still does miss, which is why it's still in beta status and not fully in production as with our other APIs. Would love to hear about your use case.
16 comments
[ 5.4 ms ] story [ 43.5 ms ] threadClassifying individual parts of pages (as Diffbot seems to be doing) is difficult, but I suspect google could take screenshots of pages reported as spam or whatever as one class and compare those to screenshots of pages w/high pr to get a pretty interesting classifier they could use as an extra datapoint. Could be an interesting experiment anyway, using data they've got lying around.
Beyond that, it's useful for companies that perform analytics or use the information to do better search or auto-categorization.
Keep it up!
How does caching works? Is there any focus on security? Multiple geolocations?
I liked the TOS :) ---- Diffbot.com is made available for personal, non-commercial, and commercial purposes. Services are provided as-is, and we do not make any guarantees on the quality or performance.
The page tagging technology looks good though.