This paper introduces a novel evolving fuzzy classifier that begins with no initial structure and develops incrementally through a participatory learning-based clustering algorithm. It employs multivariable Gaussian membership functions for rule antecedents and class outputs for consequents. The classifier’s learning algorithm is designed to adjust dynamically by creating, merging, deleting, and updating clusters and rules. Uniquely, it features a ‘procrastination’ approach where clusters are initially formed in a disabled state to robustly manage outliers and ensure only representative data influence the model. Clusters are refined based on compatibility measures using the Mahalanobis distance, with adjustments to learning rates influenced by the nature of incoming data-slowing for anomalies and accelerating for typical inputs. This mechanism enhances adaptability and model accuracy, distinguishing it from existing fuzzy classifiers. Comparative analyses on binary and multiclass tasks demonstrate superior or competitive performance, underscoring the classifier’s innovative approach to evolving fuzzy classification.
eFCMG: an evolving fuzzy classifier with participatory learning and multivariable gaussian for data stream
Paulo Vitor Campos SouzaMembro del Collaboration Group
2025-01-01
Abstract
This paper introduces a novel evolving fuzzy classifier that begins with no initial structure and develops incrementally through a participatory learning-based clustering algorithm. It employs multivariable Gaussian membership functions for rule antecedents and class outputs for consequents. The classifier’s learning algorithm is designed to adjust dynamically by creating, merging, deleting, and updating clusters and rules. Uniquely, it features a ‘procrastination’ approach where clusters are initially formed in a disabled state to robustly manage outliers and ensure only representative data influence the model. Clusters are refined based on compatibility measures using the Mahalanobis distance, with adjustments to learning rates influenced by the nature of incoming data-slowing for anomalies and accelerating for typical inputs. This mechanism enhances adaptability and model accuracy, distinguishing it from existing fuzzy classifiers. Comparative analyses on binary and multiclass tasks demonstrate superior or competitive performance, underscoring the classifier’s innovative approach to evolving fuzzy classification.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.