Google Ad Click Prediction for Humans(leve.rs)
leve.rs
Google Ad Click Prediction for Humans
http://leve.rs/blog/google-ad-click-prediction-humans/
3 comments
Spot on, about caring about conversions, not the click through rate.
Also true that this seems like a nuclear grade weapon. We have a system at optimine that is similar in a lot of ways, but also a lot less complex. Of course we are representing the advertiser side of the equation so, completely different design goals.
Also true that this seems like a nuclear grade weapon. We have a system at optimine that is similar in a lot of ways, but also a lot less complex. Of course we are representing the advertiser side of the equation so, completely different design goals.
I wouldn't classify this as "nuclear grade", not in the least. I've seen the gentleman in the cubicle next to mine use more sophisticated simulations than this to model whether there will be fresh coffee in the pot when he gets to work.
In fact, I was surprised at the simplicity of the technique they demonstrate. Logistic regression is a very powerful method, but it is generally chosen because it is simple to implement and both fast to train and reliable to train (there are no issues with whether or when it will converge).
Of course, as is almost always true of machine learning, knowing how the mechanism works is almost completely unhelpful to duplicate the results. Although this technique is does some pretty innovative things, it would be pretty trivial to achieve the same quality results if you didn't know their technique, but you did know what the feature set was (including any preprocessing and cleanup done on features). However, this system seems designed to give good enough results with extremely agressive performance SLAs, which I suspect is very nontrivial indeed.
In fact, I was surprised at the simplicity of the technique they demonstrate. Logistic regression is a very powerful method, but it is generally chosen because it is simple to implement and both fast to train and reliable to train (there are no issues with whether or when it will converge).
Of course, as is almost always true of machine learning, knowing how the mechanism works is almost completely unhelpful to duplicate the results. Although this technique is does some pretty innovative things, it would be pretty trivial to achieve the same quality results if you didn't know their technique, but you did know what the feature set was (including any preprocessing and cleanup done on features). However, this system seems designed to give good enough results with extremely agressive performance SLAs, which I suspect is very nontrivial indeed.
Looks like you spilled your Alphabetti Spaghetti on your blog post there
http://www.datasciencecentral.com/profiles/blogs/marrying-co...