A Novel Boolean Kernels Family for Categorical Data.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Kernel-based classifiers like Support Vector Machines (SVM) are powerful but often function as "black-box" models, lacking interpretability.
• Interpretable machine learning is crucial for understanding decision-making processes and building trust in AI systems.

Purpose of the Study:

• To propose a new family of Boolean kernels designed for categorical data.
• To enhance the interpretability of kernel-based classification models.
• To facilitate the extraction of human-readable rules from the model's embedding space.

Main Methods:

• Developed a novel family of Boolean kernels utilizing propositional formulas as features.
• Applied these Boolean kernels to categorical datasets, including artificial and benchmark examples.
• Compared the performance of Boolean kernels against established kernels like Radial Basis Function (RBF).

Main Results:

• The proposed Boolean kernels demonstrate effectiveness in classification tasks.
• Boolean kernels achieve competitive classification accuracy compared to standard kernels.
• The human-readable features derived from Boolean kernels aid in rule extraction.

Conclusions:

• Boolean kernels offer a promising approach to creating interpretable kernel-based classifiers for categorical data.
• This method balances classification performance with enhanced model transparency.
• The findings suggest a potential shift towards more interpretable machine learning models.