An Explainable Classifier based on Genetically Evolved Graph Structures

João Roberto Bertini Junior, Alberto Cano

Abstract: Trusting an algorithmic decision is much easier if it is understood how it was achieved. Therefore, data mining algorithms with explainable abilities are preferred over complex models for critical applications. Rule-based algorithms are among the easiest data classification models to understand. However, as most interpretable models, rule-based algorithms do not provide the highest accuracy. The Attribute-based Decision Graph (AbDG) is a model to represent a labeled data set as a weighted graph over the attributes. When associated with a proper algorithm, AbDGs can be used for supervised data mining tasks. An important aspect of AbDGs is that the graph encompasses the original attribute values and their relationships, which makes it easily interpretable by extracting rules. The AbDG drawback is defining a suitable graph structure for a given data set. In previous works, the authors proposed GA-AbDG, a genetic algorithm to search for an AbDG by evolving its edge set, keeping the vertex set fixed. In this paper, we propose an evolutionary algorithm and genetic operators to evolve both, vertex and edge sets, enhancing the search space of possible AbDG structures. Moreover, we associate a rule-based classifier with the AbDG to achieve explainable results. Experimental results show the proposed method outperforms GA-AbDG and five other classical interpretable classification algorithms.