For over 5 years, ROIVENUE has been the leading marketing attribution tool spreading from Prague like wildfire to many European markets and the US. Even though our platform has been recognized with numerous awards, we always strive for better and more precise solutions.
Having gathered feedback from numerous clients and billions of data points, we set out to build an even smarter and more nimble attribution model than we have used so far. After working on the model for a couple years, and consequently running the old and new in parallel for a long time, we are introducing the all new data-driven attribution model based on artificial intelligence and Recurrent Neural Networks (RNN). It will replace the currently used classical Markov chains and Shapley value based models.
Why data-driven attribution?
When spending thousands of euros in online marketing every week, each marketer wants to get the most out of their investment. Precise evaluation of ROI on each channel is a must in that case. Even the best analysis is pointless when the source data is off, and that’s where attribution modeling comes in. While the last-touch model is still the most prevalent ; in recent years, more and more marketers are looking for an alternative.
For them, ROIVENUE offered a choice from three data-driven alternatives: Shapley model, Markov 1st order and Markov 2nd order models. While each of these models was a step in the right direction in looking for the real ROI, they are still statistical models and as such, had a few downsides when compared with the new ROIVENUE RNN Model .
What’s new and why should you switch?
While our first generation statistical models were a great step forward for most marketers, they were not completely without flaws.
First, a certain number of conversions from each channel is necessary for the models to be precise enough. That is a limitation for smaller e-commerce projects. It made it harder for them to take advantage of the models. Also, it complicates life for those who wanted attribution to be calculated not only per channel, but also at the more detailed campaign level.
However, what was missed the most by our clients was the lack of freshly attributed data being available on a daily basis. While weekly updates were enough a few years ago, more and more companies are optimizing their campaigns every day, and they need a reliable platform to turn to when it comes to this performance evaluation.
Last but not least, Markov and Shapley model’s are not adjustable enough to suit each client perfectly. They are more of a one size fits all solution, and this enables some parasitic channels to be able to sneak by and make their way into the marketing mix.
All the insufficiencies mentioned above are solved with the new ROIVENUE RNN model while maintaining all of the benefits of data-driven attribution.
Because your attribution will now be computed on the level of every touchpoint you will be able to see attributed results sliced across any dimension within your data sets, be that: channel, campaign, AdGroup; or, for instance, device category.
You can see the evolution of the models expressed in the comparison table below.
How the heck does it work and why should you trust it
Our new RNN Model is based on predicting the user’s probability of making a conversion based on their behavior. You can imagine it as having your own superhero to find patterns in thousands and thousands of different customer paths. His help would be a three part process:
- First, he will objectively observe the behavior of your customers and see each touch point. He remembers the important patterns for later.
- After he sees enough, he is able to make judgments about customer behavior. When a potential customer approaches you from channel X your hero can say how likely he is to convert. When this customer comes again but this time from channel Y, your hero can say how the likelihood of conversion changed.
- When your hero is then given already finished customer paths, he can easily decide which touch points increased the probability of a conversion, and which of them were unnecessary.
Lastly, because he knows that customer behavior can change over time, he repeats his learning to always have the best information to base his analysis on.
Let’s go deeper: Technical explanation
Our RNN Model is based on recurrent neural networks, where a sequence of touch points form an input sequence for the network, that it trains itself with to predict if a user will convert or not.
We work with over 20 different parameters of each touch point to ensure the best results. Some of the standard parameters are the time of the touch point, number of page views, bounces, conversions and more. In fact, the number of parameters and the complexity of looking for patterns is the reason why we use the RNN model. The model even has the ability to consider additional parameters if necessary.
So what’s going on inside?
The model needs to be trained before it can be used to predict the conversion probabilities (i.e. the probability that the particular session/customer will, or will not, convert).
How is that done?
First, we generate a feature vector from all of the touch points in the customer journey (each touch point stores all the information about the session, for instance its timestamp or conversion value). Sequences of these vectors then form an input for two-dimensional matrices. A three dimensional tensor composed from these matrices represents the information about the customer paths and it is used for training the model.
To improve the results and in order to accommodate for different times of the year when customers have different behavior, the model is trained independently for different time periods. Typically, the training period covers four weeks, but can be changed if needed. When your business circumstance changes, the model adapts effortlessly.
After the model is trained, we apply it to live data where the RNN model returns conversion probabilities for every given input touch point. We assume that every user interaction with an advertisement and with the website changes the probability of conversion.
Then we score (assign a coefficient to) each user interaction (touch point) as a change of conversion probability before and after the interaction happened.
Finally, if you want to see a performance of a certain channel, AdGroup or a group of campaigns (or whatever else it might be you want to analyze) we calculate coefficients within that selection and… voila! There is your attributed performance as close to truth as one may get in today’s digital marketing ecosystem.
Each Xt is one of the customer paths which enters the model. Many additional data points like date, bounces, page views and much more are included. ROIVENUE then assigns values to each of the touch points. Y stores the coefficients for each touch point.
In conclusion, there is nothing wrong with the Markov chains or Shapley value based attribution models. There is however, a better and more precise option available right now. The adaptability of the new RNN model is one of its main benefits along with its better identification of parasitic channels – goodbye bad affiliates! If you are still wondering if data driven attribution is for you, leave us a comment or book a demo to see it in action.