Answers are grounded in this paper and cite their sources.
Generalized Linear Models (GLMs) are a key tool in Explainable Machine Learning, since they provide an easy-to-understand explanation of their coefficients and the impact of the different features in predictions. However, not all accurate GLMs may be equally appealing due to presence of secondary objectives such as sparsity or other measurement costs. For this reason, several methods have been proposed in the literature to address the multiobjective problem of joint optimization of accuracy and other secondary criteria, sparsity being the most popular objective if explainability is pursued. In this work we explore how different GLMs with good accuracy can be. We consider a biobjective P -Dispersion Problem, in which P models are sought jointly maximizing an overall accuracy measure as well as a dispersion measure. Several metrics are used to measure dispersion, such as a function of the distance between their parameters or the set of features used to predict. Moreover, different scenarios for the feasible region are considered, e.g., whether the P models are to be chosen from, or added to, a list of pre-selected GLMs. All these variants are expressed as biobjective nonlinear mixed integer optimization problems. Numerical tests are also presented, showing the effectiveness and the relevance of the proposed methodology.
No takes yet. Share an insight, caveat, or question.
Carrizosa et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: