Predictors and outcomes of an adaptive model

Configuring adaptive models involves selecting potential predictors and setting outcomes that identify positive and negative customer behavior. Unless you are a highly experienced data scientist, it is strongly recommended to leave the advanced settings at their default.

The input fields you select as predictor data for an adaptive model play a crucial role in the predictive performance of that model. A model's predictive power is at its highest when you include as much relevant, yet uncorrelated, information as possible. In Pega, it is possible to make a wide set of candidate predictors available, as many as several hundred or more.

Adaptive Decision Manager (ADM) automatically selects the best subset of predictors. ADM groups predictors into sets of correlated predictors and then selects the best predictor from each group, that is, the predictor that has the strongest relationship to the outcome. In adaptive decisioning, this predictor selection process repeats periodically.

You can use several data types in adaptive analytics, including:

Numeric data - Basic numeric data such as age, income, and customer lifetime value can be used without any preprocessing. Your model automatically divides that data into relevant value ranges by dynamically defining the bin boundaries.

Symbolic data - You can feed predictors with up to 200 distinct string values without any preprocessing. Such data is automatically categorized into relevant value groups, such as the PreviousWebpage predictor in the following example. For predictors with more than 200 distinct values, group the data into fewer categories for better model performance.

Customer identifiers - Customer identifiers are symbolic or numeric variables that have a unique value for each customer. Typically, they are not useful as predictors, although they might be predictive in special cases. For example, customer identifiers that are handed out sequentially might be predictive in a churn model, as they correlate to tenure.

Codes - For meaningful numeric fields, feed code fragments to the model as separate predictors. Simple values require only basic transformation. For example, you can shorten postal codes to the first 2 or 3 characters which, in most countries, denote geographical location.

Dates - Avoid using absolute date/time values as predictors. Instead, take the time span until now (for example, derive age from the DateOfBirth field), or the time difference between various pairs of dates in your data fields (such as the DurationLastSubscription field). Additionally, you can improve predictor performance by extracting fields that denote a specific time of day, week, or month.

Text - Do not use plain text to create predictors without any preprocessing; it contains too many unique values. Instead, extract values such as intent, topic and sentiment to use as predictors. Pega features a Text Analyzer rule for this purpose.

Event streams - Do not use event streams as predictors without preprocessing, aggregate the data instead. Pega features event strategies for this purpose. As an example, this event strategy detects dropped calls.

First, (1) it listens to a real-time dataset; then (2) it filters out dropped customer calls; next (3) it stores the terminated calls for one day; (4) it counts the number of terminated calls within the one-day timeframe; and (5) it creates an event if three calls are terminated within the one-day timeframe; lastly, (6) it emits the event. The aggregates can be stored and used like any other symbolic or numeric field.

Interaction History - Past interactions are usually very predictive. You can use the Interaction History (IH) to extract fields such as the number of recent purchases, the time since last purchase, and so on. To summarize and preprocess IH data for predictions, use IH summaries. Several predictors based on IH summaries are enabled by default (and require no additional setup) for all new adaptive models. These are the group that was referenced in the last interaction, the number of days since the last interaction, and the total number of interactions.

Multidimensional data - For models that inform the initial customer decision, things such as lists of products, activities, and transaction outcomes are useful sources of information for predictors. Use your intuition and data science insight to determine the possibly relevant derivatives, for example, number-of-products, average-sentiment-last-30-days, and so on.

Interaction context - To increase the efficiency and performance of your models, do not limit the data to customer data alone. By supplementing decision process data with the interaction context, you can adjust the predictions for a customer and provide different outcomes depending on their context. Contextual data might include the reason for a call, or the way the customer uses the website or mobile app to interact with the company, etc.

Customer behavior and usage - Customer behavior and interactions, such as financial transactions, claims, calls, and complaints, are typically transactional in nature. From an adaptive analytics perspective, you can use that data to create derived fields that summarize or aggregate this data for better predictions. Examples of this type of data include average length of a call, average gigabyte usage last month, and the increase or decrease in usage over the last month compared to previous months.

Model scores - Scores from predictive models for different but related outcomes as well as other data science output might be predictive as well. If you decide to use scores as predictors in your models, evaluate whether the models that include such a score perform better at the model level by verifying the area under the curve (AUC) and success rate metrics.

In summary, to achieve the best results, use predictors that provide data from many different sources, including:

Customer profile data such as age, income, gender, and current product subscriptions. This information is usually part of the Customer Analytic Record (CAR) and is refreshed regularly.

Interaction context data such as recent web browsing information, call reasons, or input that is gathered during a conversation with the customer. This information can be highly relevant and, therefore, very predictive.

Customer behavior data such as product usage or transaction history. The strongest predictors of future behavior typically contain data about past behavior.

Model scores, which are scores derived from the off-line execution of external models.

The responses that indicate positive or negative behavior must be identified. When predicting the click-through rate for a web banner, the default value for positive behavior is Clicked and the default value for negative behavior is NoResponse.

Applications may use different words to identify positive or negative behavior, for example, Accepted may be identified as positive behavior and Rejected may be identified as negative behavior. You can add these values when needed.