I turned my project into a tool you can use to check if a domain name looks spammy. I’m not making any guarantees with it though. It could be a nifty tool if you want to have a high-quality domain name. Or at least, one which is not obviously spammy:
UPDATE: There were a few bugs in the earlier version, I’ve fixed those (but not the other ones I don’t know about yet), so now things should be (at least slightly) better!
You probably understand the basic idea of spam detection:
- The engines (and surfers) don’t like the spam pages.
- Enter the PhD-types with their fancy models, support from the engines with their massive data-centers, funding for advanced research, and whole lot more smarts than I’ve got.
To measure these things it’s interesting to look at the “true positive rate” and the “false positive rate”. The perfect algorithm has a true positive rate of 100% (or 1.0) (all spam is identified) and a false positive rate of 0% (no non-spam pages are marked as spam). However, just as there’s no such thing as both calorie-free and delicious chocolate (PLEASE tell me I’m wrong!), there is rarely a perfect algorithm. So we are faced with trade-offs.
On the one hand you could label everything as non-spam and you would never have a false positive. This is the point on the graph where x=0 and y=0. On the other hand, you could just label everything as spam and no SERPs would contain any spam pages. Of course, there would be no pages in any SERP at all, and this whole industry would have to go to back-up plans. I’d be a roadie for my wife’s Rock Band tours.
Here is a plot illustrating the trade-offs, as adapted by me from the AIRWeb 2008 conference, including my own fancy algorithm. “y=x” is the base-line random classifier (you better be smarter than this one!)
see the actual results
Clearly, the graph shows that I totally rocked 🙂 Ignore for a minute the line labeled “SEOmoz – fair”, we’ll come back to this. As you can see, at a false positive rate of 10% (0.10), I was able to successfully label over 50% of the spam pages, outperforming the worst algorithm from the workshop (Skvortsov at ~39%), and performing nearly as well as the best (Geng, et al. at ~55%). My own algorithm, SpamDilettante® (patent pending!) , developed in just two days, with only the help of our very own totally amazing intern Danny Dover and secret dev weapon Ben Hendrickson has outperformed some of the best and brightest researchers in the field of adversarial information retrieval.
Well, graphs lie. And so do I. Let me explain what’s going on here. First of all, my algorithm really does classify spam. And I really did develop it in just two days without using a link-graph, extracting complex custom web features, or racking up many days or months of compute cluster time. But there are some important caveats I’ll get to, and these are illustrated by the much worse line called “fair” above.
What I did was begin with not one of Rand’s most popular blog posts. However, this post is actually filled with excellent content (see the graph above). Most of the things I couldn’t actually compute very easily:
- High ratio of ad blocks to content
- Small amounts of unique content
- Very few direct visits
- Less likely to have links from trusted sources
- Unlikely to register with Google/Yahoo!/MSN Local Services
- Many levels of links away from highly trusted websites
- Cloaking based on user-agent or IP address is common
The list goes on. These are not things that even some of the researchers backed by the engines could get a hold of. So I ignored these. After all, how important could knowing if there’s traffic to a site or cloaking going on? The big guys don’t care about that, right?
However, some of the things I could get just from the domain name:
- Long domain names
- .info, .cc, .us and other cheap, easy to grab TLDs
- Use of common, high-commercial value spam keywords in the domain name
- More likely to contain multiple hyphens in the domain name
- Less likely to have .com or .org extensions
- Almost never have .mil, .edu or .gov extensions
So I figured a linear regression (a very simple statistical model) based on these factors would be a pretty neat brief project. So I grabbed Danny and got some of the other mozzers to put together a rather short list of somewhat random (but valid) domain names. Danny spent an afternoon browsing all kinds of filth that I’m sure his professors (and family) are pleased he’s seeing during his employment at SEOmoz. In the end we had about 1/3 of our set labeled as spam, 1/3 labeled as non-spam, and 1/3 labeled as “unknown” (mostly non-english sites which probably have great content).
With the hard work done for me, I wrote a script to extract the above features (in just a few lines of python code). I took the 2/3 of the data which was labeled as spam/non-spam and divided it into an 80% “training set” and 20% “test set”. This is important because if you can see the labels for all your data you might as well just hard-code what’s spam and what’s not. And then you win (remember that “perfect” algorithm?). Anyway, I just did a linear regression on this 80% set and got my classifier.
To get performance numbers I used my classifier on my reserved 20% test set. Basically it spewed a bunch of numbers like “0.87655” which you could think of as probabilities. To get the above curve, I tried a series of thresholds (e.g. IF prob > 0.7 THEN spam ELSE not spam). This is the trade-off between false positive and false negative, and gives me the above curves.
And that’s the story of how I beat the academicians.
O.K., back to reality for a moment; on to the caveats.
- As I pointed out in the introduction to the competition my data set is a much simpler classification problem (complete label coverage and almost no class imbalance)
- As Rebecca says, “it’s just one of those “common seo knowledge” things–.info, .biz, a lot of .net [are spam],” and my dataset includes lot of these “easy targets”. The competition is all .uk (and mostly co.uk)
- My dataset is awfully small and likely has all kinds of sampling problems. My results probably do not generalize.
But these are just guesses. Can we support the hypothesis that I do not, in fact, rock? Well that’s what the “fair” line above is all about. I actually downloaded the data set from the challenge (jus the urls and labels) and ran my classifier on it. Suddenly my competitive system doesn’t look so good. I had a professor that had a rubric for these things, and according to his rubric I’m just one notch above “terrible” stuck squarely at “poor”. The next closest guy in the competition is “good” (skipping “mediocre”, of course) and the best system is “excellent”.
As E. Garcia said in his original post which started me on this, “it is time to revisit [the] drawing board.”
