Selection Policies

What are Selection Policies?

Conductrics agents operate under a set of high level policies that determine how it will make selections for users.

The selection policy is returned as part of the API response when using our REST API or Local JS API. They are also included in your page's Data Layer if you have our Data Layer Support enabled.

Here are the selection policies you may encounter:

Policy	Definition	Code
Random	The selection was made based on the random policy of the agent. This is the AB Testing / MVT policy. By default it will be a uniform random (each option has the same chance to be selected), but that can be adjusted by the client by using the probability weights . This policy is usually what one thinks of when performing experiments / AB testing.	`r`
Fixed	For when you want to control what drives the assignment using a rule you specify. For instance, you might have a targeting rule which says that people in the "Gold" loyalty tier should always get the "C" Variation. Any time this rule is fired, Conductrics will store the selection and subsequent conversion under the 'Fixed' policy.	`f`
Predictive	The selection was made by Conductrics' machine learning algorithm. By default, Conductrics uses a version of Thompson sampling/Posterior distribution draws based on the predicted value of each option for each of the generated audiences. See also How Predictive Agents Work and our Audience Report.	`a`
Paused	Indicates that the selected variation was returned because the agent is "paused" at that variation. These selections are NOT counted in the Conductrics reporting. This policy is also used if the visitor has been flagged as a "Do Not Track" visitor, due to browser preferences. See note below.	`p`
(none)	In some cases, the agent may not make a selection at all (for instance, if the agent is in a Mutually Exclusive relationship and is not "assigned" to the visitor.	`x`

📘
The selection policy is returned when getting an agent selection via the Local JS API and the API Usage via Web Service. It also appears in the Data Layer in you are using out Data Layer Support.

By separating out the data by each policy, we can isolate the effects from random testing, Conductrics' machine learning predictions, and the client's domain knowledge encoded in the fixed rules. If these results were all pooled together it would not be possible to separate the performance of each decision option from the effects of targeting (from both the predictive and fixed client rules).

🚧
Conductrics will treat visitors that have a "Do Not Track" option set in their browser preferences as if the agent was "paused". Visitors will be assigned to the "default" variation for each agent (the default variation is "A" unless the agent is actually paused, in which case the "paused at" variation becomes the default). These variation assignments are not counted in the Conductrics reporting.

A Note on Predictive Policies

By default, even when an agent has been set to make predictive selections, some users will still be treated under the random policy. There are two primary reasons for this:

To ensure that the predictive model is able to discover changes in user behavior over time.
To provide a random baseline to test the actual effectiveness of Conductrics' predictive selections.

One way to think about the predictive policy, is to consider it as just another option, or treatment alternative in your AB Test. For example, say we have two options, 'Offer-A' and 'Offer-B'. We can think about each one of them as a type of simple policy, or rule - '{Pick Offer-A}' or '{Pick Offer-B}'. However, predictive targeting is really just another rule, but it is one that considers aspects of the user and the context. For example '{If User is Rural & Repeat Then Pick Offer-A; Else Pick Offer-B}'.

A standard A/B test would just let you compare 'Pick Offer-A' vs 'Pick Offer-B'. Conductrics adds to this the ability to evaluate the predictive policy alongside each of the simple rules. This is possible because some of the users, even when the agent is set to predictive, are still randomly assigned selections (see A/B Testing Report for more information).