We use anonymised test and train sets from our customers’ historic data to build and validate our models.
These sets are made up of thousands of transacting and non-transacting journeys.
The transacting journeys have transaction events removed.
Given these user journeys; (impressions, click throughs and visit hits) we create a vector-based model that predicts whether a transaction would happen or not.
This is then compared to the complete journeys, including transactions, to test how accurate the model is at predicting transactions, and hence judging the engagement of an individual user at different points in their user journey.
Given that the model can weigh each event on how important it is for the transaction or not. That weighting would be used as part of the attributed share.
The model is only used if it has an accuracy of 85%+.