Predictive Analytics Director
- provides a data analyst with a model factory not a data mining laboratory
- automates best practices for fast model development.
- have a number of templates which significantly simplify a creation of new models.
advantages of the Pega Decision Management models
we save them directly from the Predictive Analytics Director portal as a predictive model rule for future use in decision strategies and process flows.
Predictive Modelling Steps
•Step 1. Data preparation
This is where you identify the decision requiring a predictive model, select the input data, and define the behaviour you want to predict.
•Step 2. Data analysis
In this step you prepare the data and develop relationships between potential predictor groupings and the outcome to be predicted.
•Step 3. Model development
Next you analyse how the predictors work together and then create predictive models using Regression and Decision Trees.
•Step 4. Model analysis
After building our models you can compare them with each other to assess their performance. Typically you will look at their behaviour prediction capabilities and how they segment customers into classes according to predicted behaviour.
•Step 5. Model export
In the final step of model development you will generate the model, including the definition of which fields the model should output. You can also make the model available to Decision Strategy Manager by saving it as a predictive model rule configuration.
Scoring models
- scores are calculated ( numerical scale).
- high scores are associated with better business (“good” behavior) and low scores are associated with worse (“bad” behavior).
- broken into intervals of increasing likelihood of one of the two types of behaviour.
- behaviors have to be classified into two distinct forms like positive and negative.
Examples:
•Responding to a mailing or not
•Repaying loans or going into arrears
•Making an insurance claim or not
You can also create extended scoring models which can include cases where the behaviour is unknown. For example, someone who has been refused a loan cannot repay it or go into arrears.
Spectrum Models
- extend the ideas of scoring models to the prediction of continuous behaviour.
The score calculated for each case places it on a spectrum from the lowest value to the highest value. Again, the score range is broken into intervals. Associated with each interval is the average value of the development sample cases that fall into the interval. These average values provide the predicted value for new cases falling into each interval.
Typical applications of Spectrum models include the prediction of:
•Likely purchase value of responders to a direct mail campaign
•The likely eventual write-off of cases recently falling into arrears
•The likely claim value of health, motor or contents insurance policy holders
A spectrum model allows you to differentiate between good, better and best.
What is Coefficient of Concordance?
A very important aspect of each model is its performance, i.e. how good is a model or a given predictor in predicting the required behaviour. We use a fancy term “Coefficient of Concordance” or “CoC” as the measure of the performance of predictors and models. You could describe CoC as a measure of how good the model is in discriminating between good cases from bad cases. The value of CoC ranges between 50% a random distribution and 100 % the perfect discrimination.
Model export report contains …
- A project summary
- A visualisation of the whole Decision Tree
- The sensitivity of the model for each of the input fields
- How the model is segmented
- Detailed insight in the analysis, grouping and validation of each of the attributes
- When the model was developed and by whom
PMML supported model types in Pega
TreeModel SupportVectorMachineModel Scorecard finalSetModel RegressionModel NeuralNetwork NearestNeighborModel NaiveBayesModel MiningModel GeneralRegressionModel
Model output?
The
output of the model will be mapped to the pxSegment strategy property when you reference the model
in a decision strategy.
How to measure performance?
Neither Pega’s predictive models nor PMML models output the runtime performance of an individual
model. To achieve this you need to create a feedback loop to compare the prediction of a model against
the actual user response. Pega Decision Management supports a very neat pattern which enables you
to do just that. Pega Adaptive Models predict and learn in real time while continuously reporting on
their performance. In this pattern we will use an adaptive model to monitor the performance of a
predictive model. If we have more than one predictive model we can then compare their respective
prediction accuracy.
Adaptive Decision Manager/Adaptive Models
- allows the business to build self-learning,
adaptive models and continuously improve predictions about customer behavior. - automatically detects changes and
acts on them in real time, enabling business processes and customer interactions to be instantly
adapted to account for changing customer interests and needs. - learns about customer behavior in real-time.
- increasingly accurate decisions are
made automatically by adapting models after each response to a proposition. - adaptive decisioning can calculate who is likely to accept/reject an offer
without any historical information. - captures and analyses data to deliver
predictions where no history is available to develop models offline. - particularly useful where the
behavior itself is volatile. - in cases where data is available for offline modelling, predictive models can be
used as an alternative, or in conjunction. - creates on the fly scoring models which are used for predictions.
- business users define the configuration of
adaptive models using the adaptive model record. - models themselves are created automatically
when a strategy using them is executed. - must be defined on the customer class in the Decision category.
The full adaptive
modelling cycle comprises the following steps:
Capture data real-time from every customer interaction.
Regularly:
o Use sophisticated auto-grouping to create coarse-grained, statistically reliable numeric intervals, or sets of symbols.
o Use predictor grouping to assess inter-correlations in the data.
o Use predictor’s selection to establish an uncorrelated view that contains all relevant aspects to the proposition.
o Use the resulting statistically robust adaptive scoring model for scoring customers.
Whenever new data is available, update the scoring model.
Adaptive decisioning
- can build channel specific models to account for differences in customer behavior when responding to outbound offers as against behavior when responding to the real-time inbound offers.
- models are directly actionable from strategies and can be used as a powerful tool in predicting customer behavior. A typical use case where adaptive models are of particular interest is in detecting complex fraud patterns in real time or predicting customers behavior following an introduction of a new offering.
Adaptive model outputs:
Propensity,
Performance and Evidence.
Propensity – The predicted likelihood of positive behavior. For example, the likelihood of a customer
accepting an offer. The propensity will start at 0.5 or 50% because at the beginning we have no
information on which to base our predictions.
Performance - How good is the model in differentiating between positive and negative behavior. Again
the initial value for the performance is 0.5, similar to chances when flipping a coin. Performance of 1.0
is a perfect prediction, always correct. Therefore the Performance should be somewhere in between 0.5
and 1.0. We would generally use performance to differentiate between two models relating to the same
proposition.
Evidence - The number of customers historically assessed by this model and who exhibited statistically
similar behavior when responding to the offer being evaluated. This is not the same as the number of
responses to a relevant proposition.
In strategies the model propensity is mapped automatically to the strategy property called pyPropensity.
There is no automatic mapping for the Performance or Evidence properties. This can be optionally set to
any strategy properties on the “Output mapping” tab.
