Why Wikipedia's A/B testing is wrong

(Synference cofounder here)

That's a good blog post.

There are many ways to handle multi-variate data, with different tradeoffs. The linear model you describe in your post is one of the classic machine learning models often used. In the post, we address the shortcoming of such models in the subsection "Trading off exploration vs. Exploitation."

As we point out, such a model ignores the trade-off between exploration and exploitation. That means that you have to gather all your data by randomly serving variants, before you can figure out the right variant to serve to each user.

That's very expensive in terms of serving-the-wrong-variant, if what you are trying to predict is conversion or revenue. By contrast, our method will quickly realise if a particular variant has a very high expected value for the current user (with the current user's particular combination of features), and so serve that user that variant - even if we aren't confident about other users.

So, like any bandit algorithm, we trade off exploration and exploitation; but further, we trade this off not just across the population as a whole, but on a personalised basis. That's a very important functionality that the approach you describe doesn't have.

Why is this an issue? Because if you have to gather data by randomly sampling, what if a certain sort of user doesn't appear very often? (e.g. lets say for a particular site, that's a 65 year old female from Canada). Do you randomly sample until you see a user like this? If so, you are randomly sampling for a long time. Or do you stop your exploration before you see rare users? This totally neglects your rarer users. What if rare users have unusual preferences, or are particularly valuable? (E.g. maybe your rare users are your biggest spenders - which happens.)

And, if you have high dimensional data, all your users are effectively rare (because you rarely see users with that exact combination of features).

Furthermore, the 'gather data randomly' approach is also impractical if user behaviour is dynamic, as it often is in real-world settings. (i.e. the preference of your user base as a whole changes - maybe they've all recently read an article, so a different marketing copy now resonates with them)

Finally, sometimes those interactions you mention, between different features, are really important, and shouldn't be neglected.

I could go on - there's a lot of subtle issues that the approach we are using solves that simple multivariate testing doesn't handle, basically.

What we're doing is trying to provide an API that wraps all this knowledge up into a service that's easy for people to use.

Your blog post makes an important mistake in that it recommends actually inverting the matrix X^T X, which one never actually does in this case. In general, A^{-1} b is sort of mathematical "slang" for "a solution to the linear system Ax = b", not "compute A^{-1} and multiply b by it".

This is discussed in many places, but here are a few examples:

http://www.johndcook.com/blog/2010/01/19/dont-invert-that-ma... http://en.wikipedia.org/wiki/Linear_least_squares_(mathemati...

Perhaps this article title has been A-B tested and found to be highly successful at driving traffic to the site.

(Synference cofounder here)

It has successfully increased Wikipedia's donation revenue; however, some fundamental assumptions baked into the AB-testing approach are all wrong; this leaves a lot of money on the table.

Instead of trying to find a single best version for everyone, people should realise that different segments of their users are going to have different preferences. Machine learning can find these in an automated way, and there's a lot of value to be gained - but people need to see past the standard 'one size fits all' AB-testing approach to take advantage of it.

>The regret bounds on contextual bandits are O(d sqrt(T)), where d is the dimensionality of the user profile you're using and T is time (see e.g. http://arxiv.org/abs/1209.3352). In contrast, the standard stochastic bandit has regret bounds of O(log(T)), which is much tighter. The summary is you need a lot more data to explore the bigger range of possibilities you've created. So I'm not sure contextual bandits are viable for smaller sites, but Wikipedia definitely has the traffic to make them work.

Its definitely viable for smaller sites. Your argument there is about the bounds in a particular theoretical situation ("when the contexts are provided by an adaptive adversary.")

In a real world situation, your contexts are provided by the normal behaviour of your users, not an adversary. And, rather than the patterns being maximally hard to find, there'll instead be a lot of low-hanging-fruit - in terms of very obvious combinations of features that define different segments of users.

What's important is not the bound of the contextual bandit in a theoretical situation, but rather what performance it gives in 90% of business situations. In many situations, doing some personalisation to automatically find the most obvious user groups is a huge win.

And later, as the amount of data increases, the granularity of segmentation will effectively increase.

I agree with your general comments on bandits, and love what Myna is doing.

I'm a cofounder of the early 2-person startup whose blog post this is.

Yes, it's currently difficult and expensive to bring complex ML to bear on this problem; we are trying to make that easy and cheap. The reason people adopted AB-testing was that it has a positive impact on the metrics they care about; we think the workflow can be easier and more valuable than AB testing. If its easy to use, the extra value should cover the cost.

We're currently focused on trying to communicate that this approach can create value - this blog post, written for a technical audience, is to give some intuition around how the approach works - does that make sense?

We genuinely believe in making it easy to use ML optimisation in this context - that's why we're working on this.

That's because this post exists as attempted marketing for Synference.

Regarding the title, I've updated the title to say 'conceptually wrong' - its more of a mouthful, but maybe what we mean is clearer? I hope our intent was clear in the intro paragraph anyway.

Hey there, I wrote that post. It's a valid criticism you make, but the main point of the article was not to pick on Wikipedia, which is actually way ahead of most other organizations in terms of testing its content.

We made a case-study Wikipedia because their banners are a well-known example of A/B testing. Plus, because their banners make tens of millions of dollars each year, and they have an entire team of people working on the A/B tests over a period of months, I think they have capacity to use something more sophisticated. Maybe, like you suggest, they're just focused on other aspects of the fundraising campaign. I guess the only way to know would be to hear from some of the Wikipedia team!

You could divide articles into various categories.

Maybe people reading computer science articles prefer light blue, while people reading medicine articles prefer gold.

Of course, a quick look at Wikipedia's categories show there are many possible article categorisation schemes (are sneakers footwear like stilettos or sports equipment like cricket bats? Of course they're both), so this would introduce many additional variables - more than you could exhaustively A/B test.

I'm one of the Synference co-founders, and we've thought a lot about how to make it easy to extract relevant information from users who land on your site for the first time.

There are many important pieces of information you can grab from a user who's never been on your site based on basic information available to all web servers, such as the user's IP address and user agent: (1) the time of day/day of week they're visiting -- relevant because, e.g., night users or weekend users might be a different demographic (2) information related to the particular page they're visiting, such as article category on wikipedia (3) operating system type and age (4) device type -- smartphone, tablet, pc (5) geographic location.

You can use location to infer a range of demographic features, such as income, education, etc. Due to its importance, location is an attribute that Synference adaptively coarsens/zooms in on. For example, if you have a lot of users from a certain city, then each user within that city will have a location whose granularity goes down to the neighborhood. But if you have relatively few users from Europe, then each user's location might only be at the level of the country.

The Synference API makes it easy to extract this information. All you have to do is send the API the user's IP address and user agent, and the API parses those two pieces of information into most of the features I've just listed. We also add a timestamp (corrected for the user's timezone using the IP address's geographic location).

We don't need to split users into buckets---machine learning classifiers build models that automatically take care of deciding how a user's attributes can be used to predict the target metric.