The scorer determines the best option to show to a visitor in real-time.
The scorer compares the visitor's profile and behavior with the training data, scores each content option, and selects the best set of offers, or experience, to return to the visitor's page. It is important to note that 10% of traffic is set to be randomly served content. The model serves randomly to avoid becoming stuck on one option. For example, if one option began with a higher probability of conversion, the algorithm would favor that option and begin to re-enforce its decision with excessive serves of that option. To ensure the algorithm continues to consider the possibility of other options, these options are randomly explored to gauge their performance. For the other 90%, the scorer follows a series of steps to choose the best content option for each mbox in the campaign.
Scoring decisions are made at the modeling group level. A modeling group is a set of offers with similar marketing themes. Users have complete control over the modeling groups, and can determine which offers should go in which group. More detail is available in Creating an Automated Decisioning Campaign.
Two scores per modeling group are determined for each person: a personalized profile score and a generalized score. The personalized score is how the model expects each person to react to the content based on his individual variables. The generalized score is the sum of all the response rates of all offers within a modeling group for users in the random serve group. This establishes a general popularity hierarchy of modeling groups. The personalized score and generalized score are added together, and this combined score is used to determine the score for each modeling group for which the visitor is eligible. Because the generalized score includes more traffic, temporal changes are reflected much more quickly than in the personalized score. Adding them together provides the best opportunity for lift.
If a visitor has no profile values that match the variables in the model, then the generalized score is used.
Once the model runs, content is chosen for new visits based on what the model predicts these new visitors will respond to favorably. The score from the model is evaluated with a real-time reaction update model to choose the best content to deliver.
Targeting criteria must be met for each offer first, and then the modeling group with the top score is selected. If there are multiple locations in the campaign, the scorer picks the top-scored modeling group for the second location, ignoring any options that are not allowed due to experience exclusions. Through this process, a modeling group is selected for each location in the campaign. If multiple offers exist in the winning modeling group, then one of those offers is chosen randomly for delivery. Targeting criteria and experience exclusions are described in Creating an Automated Decisioning Campaign.
Sophisticated algorithms are used to determine the "best score" for a particular profile and offer set. When a campaign launches, the model does not know what is the best to serve, so each option is served an equal amount. However, the system learns very quickly, so a strictly equal split happens only at the very beginning of a new campaign. As the system's knowledge quickly increases, it begins to favor the options that indicate success. It takes input from the lower performing modeling groups as well, via the exploration traffic, to ensure that its decisions are correct. Finally, the model score is influenced by the real-time reaction tracking to reach a final score for each modeling group and mbox combination to determine the experience to deliver to each visitor.