A classification method based on dominance neighborhood granularity

Abstract

Classification algorithms, widely utilized in various practical settings, are increasingly demanded in critical domains to not only display high accuracy but also provide interpretable results. Granular Computing (GrC) theory, recognized for its capability to clarify complex data and issues through its concept of granularity, has become an intuitive and straightforward approach in problem-solving. Nevertheless, contemporary granular computing-based classification techniques are largely confined to the processing of Euclidean features of data, thereby neglecting relationships and structures between data within the data set. In response, we introduce a classification method based on dominance neighborhood granularity (CDNG). CDNG, while accommodating Euclidean features, also incorporates non-Euclidean features to obtain diverse information, thereby augmenting predictive accuracy. Specifically, CDNG establishes a dominance neighborhood granularity drawn from neighborhood granularity and dominance granularity. Subsequently, we appraise CDNG using real datasets and performance metrics, including accuracy (ACC), Recall Score (RS), and F1 score (F1), contrasting our findings with other classification methodologies. We conclude with an analysis of parameter experiments and multi-angle noise experiments specific to CDNG. The experimental results affirm that CDNG outperforms other methods in terms of accuracy while maintaining robustness concomitant with high accuracy.