Understanding Decision Management
|
|
Decision Management functionality provides sophisticated mechanisms to create applications that determine which processes to run, and which products to offer to customers through the Next Best Action principle. This principle increases customer loyalty by addressing multiple issues in the decision making process. Decision Management functionality includes:
Decision Management functionality is delivered through the combination of Decision Management rules sets and the Decision Management service layer for adaptive decisioning and business monitoring.
PRPC includes a sample application (DMSample) that illustrates typical use cases of Decision Management functionality without Adaptive Decision Manager and Visual Business Director. You have two ways to access this application:
Landing Pages
Access Decision Management landing pages through the Decisioning entry in the Designer Studio menu. The menu is structured by category:
Rule Types
Decision Management supplies the following rule types:
Glossary of DSM terms
The following terms are often associated with the configuration, implementation and use of the predictive analytics and Decision Management.
Predictive Analytics Director (PAD) is used to develop and generate predictive models. PAD develops the means to differentiate between cases on the basis of likely future outcome.
Powerful and reliable predictive models deliver the key insights that enable opportunities and risks to be evaluated, constituting the foundation of personalized strategies. PAD reveals the relationships in your data, the critical information, and the interactions that drive customer behavior in an intelligent data mining process that knows what needs to be done. Your role is to define the objectives and judge the results.
The main process of the Adaptive Decision Manager. The engine is responsible for storing sufficient adaptive statistics analyzing them, and producing individual scoring models that are used in PRPC. These statistics keep the relevant values for adaptive models defined in decision strategies. From these statistics, the adaptive analytics engine creates scoring models that are published to the adaptive data store. PRPC retrieves the scoring models from the database, and uses them to calculate the prediction.
The database scoring adaptive statistics and adaptive models.
Adaptive models are scoring models created through strategy execution when the strategy contains adaptive model components. These models output predictions calculated and adapted in real time as interaction results are captured. Models in the ADM system are configured through adaptive model rules that define the settings that influence the behavior of the adaptive models in the ADM system.
When adding adaptive model components to a strategy, you configure the propositions the adaptive model is going to model and the interpretation of the outputs.
Adaptive models belong to the self-learning aspect of Decision Management. They are typically used in the absence of historical records.
The persistent information resulting from running a strategy containing adaptive models.
A behavioral profile represents a model created on the basis of univariate performance. The probabilities of positive behavior for each interval/category are score bands that can be used to predict in the same way as those of any other model.
A case can be a person, company or event that exhibits some defined outcome.
A weight that is used for each predictor in the logistic regression formula. The coefficient is an indication of the importance of a predictor. Negative coefficients imply the presence of predictors with very similar behavioral profile. If present, they can lead to over fitting and unreliable models. Consider reanalyzing the predictor grouping to ensure predictors with highly correlated behavior are placed in the same predictor group.
The Coefficient of Concordance (CoC) is a non-parametric coefficient sensitive to the complete range of score bands irrespective of their distribution.The CoC measures how well the scores generated by the model separate positive from negative outcome cases using the statistic known as coefficient of concordance. CoC can vary between 50% (a random distribution of positive and negative cases by score band) and 100% (a perfect separation). The minimum is 50% because the scores are simply used in reverse if a set of scores orders negative cases before positive cases. Its virtue as a measure is that it encourages models to be predictive across the score range. If the desired operational circumstances (volume or quality of business) are unknown, CoC generates powerful and generalized models.
Contact policy rules define periodic limitations for one or more channels. Contact policy rules are in the Marketing category that is enabled in a Next-Best-Action Marketing (NBAM) application.
Data about customers, and their previous behavior. This data can be used for modeling and strategy design. A source should contain one record per customer with the same structure for each record. Ideally, data should be present for all fields and customers, but some missing data can be tolerated.
The result of running a strategy in the interaction context. Several decisions can be involved in a single interaction.
Geofence rules define the latitude and longitude of a location and the radius of the fence surrounding the location. Geofence rules are in the Marketing category that is enabled in a Next-Best-Action Marketing (NBAM) application.
Some contact with the customer in real time, or offline.
The reaction of a customer to a proposition. Interaction results are recorded in the Interaction History database tables and propagated to ADM and VBD.
Typically, the values of numeric predictors are grouped in intervals. Each interval provides a useful building block for understanding behavior.
A measure (multiplied by 100) of the improvement in behavior exhibited by cases in one interval or segment over the average of all cases.
The Next Best Action (NBA) strategy lets applications take the best decision in a multidimensional context (retention, recruitment, risk, recommendation, etc.).
Omega XML Language. The XML file format of predictive models as published using Predictive Analytics Director.
Predictive Model Markup Language. An XML-based language that provides compatibility methods for applications to define statistical and data mining models and further sharing these models between PMML compliant applications.
The group of cases with known behavior, which is consistent with the group of cases whose behavior is to be predicted. In predictive analytics, it is from the population that samples are extracted for modeling and validation.
The behavior to be predicted. The behavior is specific to a form of outcome at a given point in time.
An algorithm that delivers predicted behavior and values for one or more segments given the input of the required data about a case. Predictive models are developed in Predictive Analytics Director.
Some measure of the scores or segments generated by models. Performance can be measured in terms of predictive power, value, or rate achieved under selected conditions.
The predictive power of a scoring model is the measure of the ability of a model to separate cases with positive and negative behavior.
The grouping of predictors whose relationship with behavior are correlated at (or above) a selected level of correlation.
Predictors are properties considered to have a predictive relationship with the outcome. Predictors contain information available about the cases whose values may potentially show some association with the behavior you are trying to predict. Examples include:
The probability of positive behavior or membership.
A tangible (a handset or a subscription) or less tangible (benefits, compensations or services) product offer.
Proposition bundling is a method of combining and presenting a number of propositions as a coherent and justifiable set in terms of cross-product eligibility, propensity and likelihood of interest linked to the call reason. The proposition set is provided in a bundle (for example, the cheapest proposition is offered at a reduced price or for free, a discount is given on all propositions and there are additional free propositions).
A sub-set of historical data extracted by applying a selection and/or sampling method on the data source. To be meaningful and reliable, it is essential that sufficient records are used and that the distribution of values and patterns of behavior are representative of those in the population.
The value calculated by the model that places a case on a numerical scale. High scores are associated with good performance and low scores with bad performance. Typically, the range of scores is broken in intervals of increasing likelihood of one of the two types of behavior (positive or negative), based on the behavior of the cases in the development sample that fall into each interval. Score intervals are aggregated under a score band.
A score band is a set of score intervals.
The way a predictive model segments the cases in the population.
The scorecard decision rule calculates segments by combining a number of properties. The resulting segmentation is translated in a score.
The value calculated by the model, known as the score, places a case on a numerical scale. High scores are associated with good performance and low scores are associated with bad performance. Typically, the range of scores is broken in intervals of increasing likelihood of one of the two types of behavior (positive or negative), based on the behavior of the cases in the development sample that fall into each interval.
A group of customers defined by predicted behavior, score, and characteristics. Segments are implemented through segmentation components in a strategy. They drive the decision flow by placing a customer in a given segment for which actions/results are defined.
Segment rules define how to retrieve a population and check if a customer falls in a segment or not. Segment rules are in the Marketing category that is enabled in a Next-Best-Action Marketing (NBAM) application.
Simulations are executed based on changes in the strategy. The strategy decides the top propositions to be offered to the customer.
The reasoning built up by a set of components that allow you to define the business strategy. A strategy provides the decision support to manage the interaction in the context of the decision hierarchy. Each component has a well defined functionality. A strategy can reference other decision rules (scorecards, predictive models, decision tables, decision trees, adaptive models, and strategies), and import data and propositions.
Symbolic predictors can be treated as categorical or ordinal data. Numeric predictors can be treated as categorical or continuous data. Categorical treatment captures data based on the probabilities of positive behavior for each interval/category, ordinal treatment on the sequence code of each category and continuous treatment on the raw data of the predictor.
Detecting trends is possible by comparing the performance of multiple models that are triggered by the same propositions but are configured with different performance window size to determine the time frame in number of cases over which the performance is calculated. Implementing trend detection requires a combination of strategy design patterns and using compatible adaptive models.
Univariate performance represents the potential performance of a predictor on its own.
|
About Predictive Models
About Scorecards About Adaptive Models About Strategiess About Interactions About Interactions Understanding the Decision Management agents |