Smart Testing: A/B Test Optimized with Machine Learning

Posted by Virginia Peón on Sep 15, 2020 12:14:06 PM

Captura_Test inteligente


You have to create a notification for a new campaign and you wonder:

Will the notification work better with another photo? And with a more casual text? Will it have more conversion if I add emoticons?

We all want to be right in our decisions but sometimes it is difficult to know which the best option is. Our problem of choosing the variation that will best convert can be compared to the so-called "multi-armed bandit" problem.


In English slot machines are informally known as ‘one-armed bandits’ because of the lever (arm) with which they steal your money.


Imagine that you are in front of several slot machines and you knew that with some it is easier to earn money than with others. The difficulty is knowing which is the most profitable. How can you know which machine is the best to get the most money in the shortest possible time? You start to play, but if you pull a lever, you are not pulling another arm.


The different variations of the campaign would be like a row of slot machines, each with its own probability of producing a conversion. The problem is balancing between collecting data to find out which one is the best and using the optimal version.

If we use a classic a/b test approach, initially, we would do an exploration phase in which we would analyse which version converts best. We would send all versions to the same number of users (a small group of the total). Seeing the one that gives us the best result, we could go to the exploitation phase, that is, send the most profitable version to the rest of the users.

testab captura1


It seems like a simple choice, but the reality is that it is much more complicated since it requires statistical knowledge. With the help of artificial intelligence, with the so-called “multi-armed bandit” algorithms, the exploration phase is continuous but preference is given to the one that is most profitable for us. The exploration phase and the exploitation phase are balanced with what we obtain higher earnings.


Just like in other companies like Netflix, Google or Amazon, at indigitall we have automated this process based on the “multi-armed bandit” algorithm of “Thompson sampling”.


Through our Smart testing at indigitall we help you to get to know your clients or users better, thanks to reinforced learning, exploration and exploitation at the same time and automatically of several versions of the same campaign. Increase the chances of success in your push notifications by knowing which messages your users prefer.

Once the different versions are created, you will not have to worry about analysing which one has the highest conversion. In addition, you will have the peace of mind that it will reach more users.


Topics: indigitall news & updates, Artificial Intelligence

Get indigitall free
Download our free ebooks