Explainable artificial intelligence through graph theory by generalized social network analysis-based classifier

Area of Science:

• Machine Learning
• Data Science
• Network Science

Background:

• Social Network Analysis-based Classifier (SNAc) previously handled time-series numerical data.
• Existing methods may lack interpretability and visual representation for complex datasets.

Purpose of the Study:

• To extend SNAc for tabular data classification, including numerical and categorical features.
• To develop a visual, interpretable machine learning classifier named GSNAc.
• To demonstrate GSNAc's effectiveness and compare its performance against established classifiers.

Main Methods:

• Tabular data is converted into a network graph where samples are nodes and similarities are edges.
• A visualizable 'graph classifier model-GCM' is extracted by simplifying and enriching the network graph.
• Classification is performed by mapping test nodes into the GCM and evaluating average similarity using vectorial and topological metrics.

Main Results:

• GSNAc demonstrated superior or comparable performance against well-established machine learning classifiers on benchmark datasets.
• The method successfully transforms multidimensional tabular data into a 2D visualizable network domain.
• The classifier provides a visually comprehensible and interpretable prediction process.

Conclusions:

• GSNAc is an effective supervised visual machine learning classifier for diverse tabular data.
• The primary contribution is the transformation of data into a visual network for enhanced interpretability.
• GSNAc offers a human-comprehensible and highly visual approach to machine learning classification.