From Numeric to Granular Models: A Quest for Error and Performance Analysis.

Area of Science:

• Computational Intelligence
• Data Science
• Predictive Modeling

Background:

• Traditional numeric models aim for high accuracy but may lack descriptive power for prediction outcomes.
• Analyzing and modeling prediction errors is crucial for enhancing existing predictive models.
• Information granules offer a framework for representing uncertainty and imprecision in model outputs.

Purpose of the Study:

• To propose a novel design methodology for granular models by augmenting existing numeric models.
• To improve and quantify the prediction abilities of numeric models through error analysis.
• To develop granular models that produce comprehensive prediction outcomes as information granules.

Main Methods:

• Augmenting existing numeric models with their associated prediction error.
• Developing novel granular architectures incorporating modeling errors.
• Formulating and analyzing three different architectural developments for granular constructs.

Main Results:

• A new granular model design methodology is established.
• Granular models produce granular outcomes by aggregating numeric model outputs and error terms.
• The performance of granular constructs is evaluated using coverage and specificity criteria.

Conclusions:

• Granular models offer a more descriptive and comprehensive way to assess prediction quality compared to purely numeric models.
• The coverage and specificity of information granules effectively express the quality of prediction results.
• This approach enhances the interpretability and utility of predictive modeling through granular computing.