Explaining Support Vector Machines: A Color Based Nomogram.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Support Vector Machines (SVMs) are widely used for classification and regression.
• The interpretability of SVM models, especially with non-linear kernels, remains a challenge, often leading to their use as "black boxes."
• Lack of interpretability hinders SVM adoption in critical areas requiring accountability for model decisions.

Purpose of the Study:

• Investigate the explainability of SVMs with linear, polynomial, and Radial Basis Function (RBF) kernels.
• Determine conditions under which SVMs can be explained, providing decision-based interpretations.
• Identify scenarios where SVM interpretation is currently not feasible.

Main Methods:

• Defined explainability as the ability to derive decisions from contributions of one or two input variables.
• Conducted experiments on both simulated and real-world datasets.
• Analyzed the impact of parameter values (e.g., polynomial degree, RBF kernel width, regularization constant) on SVM explainability.

Main Results:

• SVM explainability is contingent upon specific parameter settings.
• Parameter combinations yielding similar cross-validation performance show higher explainability with lower polynomial degrees or wider RBF kernels.
• A method was developed to summarize SVM classifiers, providing exact representations for linear and low-degree polynomial kernels.

Conclusions:

• The study provides a unified visualization for SVM classifiers across different kernels.
• An indication of approximation reliability is offered for kernels beyond second-degree polynomials.
• The methodology is accessible via an R package, apps, and an illustrative movie.