Excellent prediction and scoring with modelling of non-linear dependences

The Viscovery Predictor offers unique patented capabilities for both linear and non-linear prediction and scoring. The system enables workflow-oriented prediction, scoring, and non-linearity analysis within a project environment for creating, applying and evaluating prediction and scoring models.

Viscovery Predictor key benefits

Easy creation and handling of models

• Automatic splitting into training and test data sets

• Support for full regression, stepwise regression, and logistic approximation of probabilities

• Comparison of created model variants with estimated prediction error

• Score charts, gain charts, and definition of score groups; performed with a click

Superior prediction accuracy

• Use of self-organizing maps (SOMs) for partitioning data into homogeneous groups and subsequent creation of local regression models based on smaller clusters

• Automatic optimisation of receptive fields for local regression models

• Exclusive use of statistically significant, validated white-box models, guaranteeing both efficient and reliable prediction

• Patented technology for the extraction and exploitation of non-linear relationships between variables

Advanced analytical functions

• Context-sensitive tools for statistical analysis and with administration of relevant parameters

• Determination of “predictive influence” of explaining variables on the target value

• Non-linearity diagnostics

• Reporting functions

As a software module of the Viscovery Data Mining Suite, the Viscovery Predictor also provides the suite's general functionalities and benefits.

Viscovery Predictor features and functions

Viscovery Predictor provides interfaces for common databases and can easily be linked to customer databases. The user is guided through the entire model creation process by means of precisely defined workflows for creation, evaluation, and application of a predictive model.

The patented Viscovery Predictor procedure combines non-linear SOM technology with conventional linear statistics (e.g., regression analysis, principal components analysis, correlation matrices and scatter plots). Data is sorted according to overall similarity using SOM technology, and subsequently subdivided into groups that contain only very similar objects. The behavior of these homogenous groups can be predicted far more precisely than using just one group for the entire, inhomogeneous data set.

Local regressions are used within the clusters of data, thereby improving the prediction quality considerably compared to conventional prediction methods. The set of local regressions provides a validated prediction model which finally can be applied to new data records to predict target values or to score the data records according to their estimated values. The predicted values can be used immediately in applications or can be subsequently entered into a more comprehensive segmentation model.

Various graphical views (histograms, gains charts, and score charts) and other relevant statistical values (e.g., estimated prediction error) can be displayed. By automatically splitting data into training and test data sets and testing each trained model, optimal support is available for the validation of the models. Different model variants can easily be compared easily with one another.

Viscovery Predictor is also available as a stand-alone product.