Applications of a new empirical modelling framework for balancing model interpretation and prediction accuracy through the incorporation of clusters of functionally related variables

• A recently proposed modelling framework is tested and compared: NI-SL.
• NI-SL has the ability to balance model interpretation and prediction accuracy.
• The framework is based on the selection of functionally related variables.
• Four widely different real world case studies were used in the comparison study.
• Models obtained are easily interpretable and do not compromise prediction ability.

Current classification and regression methodologies are strongly focused on maximizing prediction accuracy. Interpretation is usually relegated to a second stage, after model estimation, where its parameters and related quantities are scrutinized for relevant information regarding the process and phenomena under analysis. Network-Induced Supervised Learning (NI-SL) is a recently proposed framework that balances the goals of prediction accuracy and interpretation [1], by adopting a modelling formalism that matches more closely the dependency structure of variables in current complex systems. This framework computes interpretable features that are incorporated in the final model, which effectively constrain the predictive space to be used. However, this restriction does not compromise prediction ability, which quite often is enhanced. Both classification and regression problems can be handled. Four widely different real world datasets were used to illustrate the main features claimed for the NI-SL framework.