Gaussian Process Regression (GPR) Representation in Predictive Model Markup Language (PMML).

Area of Science:

• Machine Learning
• Data Science
• Statistical Modeling

Background:

• Predictive Model Markup Language (PMML) is an XML-based standard for data mining and predictive models.
• Previous PMML versions lacked support for probabilistic machine learning algorithms, hindering uncertainty quantification.
• Gaussian Process Regression (GPR) is a powerful probabilistic method for approximating functions and quantifying prediction uncertainty.

Purpose of the Study:

• To describe Gaussian Process Regression (GPR) models represented in the Predictive Model Markup Language (PMML).
• To introduce new features in PMML 4.3 enabling confidence bounds and distributional predictions for GPR.
• To facilitate the application of GPR in manufacturing data analytics through standardized representation.

Main Methods:

• Development of GPR model representation within the PMML framework.
• Utilization of PMML 4.3 features for confidence bounds and predictive distributions.
• Demonstration of a prototype implementation using a real-world manufacturing dataset.

Main Results:

• Successful representation of GPR models in PMML 4.3.
• Inclusion of confidence bounds and distributional predictions for enhanced GPR analysis.
• Validation of the PMML-based GPR representation through a manufacturing data analytics prototype.

Conclusions:

• PMML 4.3 now supports GPR, enabling standardized uncertainty quantification.
• The standardized representation facilitates the integration and deployment of GPR in manufacturing.
• This work lays the foundation for advanced uncertainty quantification in predictive analytics.