Training data records are the visitor data that the model uses to build predictions.
Each training data record contains a visitor profile, a snapshot of the profile when content was served, whether the visitor converted, and any revenue associated with the conversion. A model generalizes from all these specific instances of training records to predict the response probability of a previously unseen visitor to each of the content options in the campaign. Training data records are stored for both targeted and exploration content serves, which allows predictions to be made for anonymous and known visitors. For example, in a campaign with three offers, the model's job is to predict the probability that a visitor will convert on each of the three offers based on that visitor's similarities to previous visitors. If the modeling goal of the automated decisioning campaign is a revenue metric, then the model makes a value prediction rather than a binary conversion prediction. The marketer can determine the modeling goal for each campaign.
Many data points are stored per profile for the model builder. The profile variables fall into two categories: automatic and custom.