Predictive Models

Predictive models are used to predict how likely a contact is to react in the same way as the target audience, e.g. to find contacts who are likely to become GPS customers we can identify those people that already are and use them as a target. Demographics can then be fed in for use by the model prediction process to identify the key characteristics of a GPS customer, and how they correlate to the rest of the domain.

Design View

  • Analysis Table - the contact or person level of the database at which the model is being produced.

  • Modeling Method - this is the algorithm and method to be used for the predictive model.

  • Target Query - this area is used to add the query for use within the model.

  • Fields - this area is used to add the Fields for use within the model.

  • Model Query - this optional query is used for specifying the contact data set for which the model should be based upon. This can contain a sample query for models to apply to the entire contact table and then removed once tested.

Once the model design is complete, the training process runs through the prediction tests using a cutting strategy to cut the data and compare variables against the target to predict a model to score. This process repeats several times and derives a score, which is accompanied by a robustness to indicate how robust the model is.

Training a Model

To train a model design, select the Train Model icon from the hamburger menu as shown below:

Once the training process is complete, which can last a few minutes for large data selections, the robustness score and statistics will be calculated.

The scoring process produces a Boolean value of 0-1 with a staged increase in likelihood up to 1. The model can then be used to determine people who are predicted to be GPS customers using a simple selection like this, where 0 means unlikely, but up to 1 is highly likely.

A 5-star model is the best possible, so we can Score this model enabling it to be used in Selections and Campaigns.

Scoring a Model

To score a model design, select the Score Model icon from the hamburger menu as shown below:

Alternatively, the Train and Score Model option can be used for re-training models that have previously been trained.

Using Predictive Models in Selections

The Model is used within a Selection using the Prediction node. The selection below is requesting the Top 500 likely GPS customers based on the model prediction score:

Once added, the criteria available for using the model is available on the context panel to the right. The model is chosen from the list, then the 'Take Type' is specified from a range of options. These are:

  • Take First (Number of records) - Takes the first N records from the table that have been scored by the model, for a specifically sized record selection.

  • Take Last (Number of records) - Takes the last N records from the table that have been scored by the model, for a specifically sized record selection.

  • Take Over (Score) - Takes all records that are above a given score, from the decimal range between 0 to 1.

  • Take Under (Score) - Takes all records that are below a given score, from the decimal range between 0 to 1.

  • Take Equal (Score) - Takes all records that are equal to a specific score.

  • Take Not Equal (Score) - Takes all records that are not equal to a specified score.

Using Predictive Models in Campaigns

The Model is used within a Campaign using the Prediction step. The criteria for how the model is used is the same as that shown above for a selection. The incoming Audience step will be evaluated to determine if any records meet the criteria of those modeled, for example, if they already bought the product.

These records then drop into the remainder total during processing, and from the model, the 'taken records' pass through to the next stage of the campaign.

In the example above the audience contains no records that meet the model requirement, a score of over 0.2. The query split step shows us 24 records belong to Individuals from the incoming Audience that already have a GPS product.

Scheduled Training & Scoring

There are situations where a model may be applied to an extremely large set of data, and the process to train and score the model can be very intensive on the server resources available to Engage. For this reason, there is an option schedule the Training and Scoring of models under the Maintenance tab.