March 28th, 2016 | Data Science

Webcongress presentation Machine Learning for Attribution Models

The past 4th and 5th of March Juan Damia, Intellignos CEO, was invited as Keynote Speaker to Webcongress Mexico. The presentation was about how to improve your predictive skills by using Machine Learning Algorithms.

Machine Learning is driving Predictive Analytics to a completely new level due to it’s unique capabilities to improve its certainty with the increase of data. More data it doesn’t mean more problems but more predictive power.

“Machine Learning is the application of algorithms (conditional models) to data series (from small to big data) with the objetive of generating knowledge”.

IMG_0844

Machine Learning can be used for face detection, objects detection, voice recognition, identifying genetic patterns and allowing self driving cars among many others things related to artificial intelligence. However Machine Learning can do much more for you and your company right now to solve some of the current pains you are facing. I’m talking about things that can impact the way you work and make decisions.

One interesting example showed at the presentation was using Machine Learning to solve the current problem with the Attribution Models. Current Analytics platforms and adservers have the following standard attribution models:

  1. First Interaction: Assigns 100% of the conversion to the traffic source that bring the user (or browser) in the first place.
  2. Last Interaction: Assigns 100% of the conversion to the traffic source that bring the user (or browser) in the last interaction before the conversion.
  3. Linear: Assigns the conversion in equal parts to all the traffic sources that interacted with the user or browser.

The problem with these models is that they are not based on information and depends on the model you chose the “reality” the model will tells you. Some people in your company will prefer to use one model than the other depending on which one shows better results. Some areas will prefer first click, some others last click and so forth.

In every case, the information is not correct, it is just based on a non-informed decision. So probably the company will decide to use the model recommended by the person with the higher decision making power (Hippo, Highest paid person’s opinion).

The solution can be found with an Attribution Model based on an algorithm that assign each traffic source, and its position on the conversion, a weight based on its real performance. To do so you can use a Supervised Machine Learning model based on regression (we are looking the specific weight and not a discrete output). The process would be:

  1. Covariance analysis between all the variables under analysis against conversion.
  2. Once defined which variables are related with conversions we generate those as dimensions to generate covariance analysis to determine the weight of each traffic source based on the generated dimension (Male vs Female, User from Mac vs PC, etc.) considering the specific position during the interaction.
  3. As result you will have the weight per traffic source and position. That information will “feed” and adjust the algorithm that will attribute the conversion in a proper manner. Since the information is in constant change, it’s key to run the process every month or week to keep the algorithm updated.

Need more information?

Fill out my online form.

Newsletter

Join our Newsletter and receive periodical news and updates: