An Evolving Gaussian Regularized Fuzzy Classifier with Incremental Feature Selection

While significant attention has been given to adaptive model structures from evolving data streams, the challenge of incrementally selecting relevant features over time is often neglected despite its potential impact on model efficiency and interpretability. This paper presents an evolving fuzzy classification approach for numerical data named eGRFC-InFS (evolving Gaussian Regularized Fuzzy Classifier with Incremental Feature Selection). eGRFC-InFS employs a one-pass incremental learning algorithm to handle non-stationary data streams. The model is built from scratch, without retaining data, and incorporates an online feature selection procedure that dynamically activates and deactivates features based on instance means, ensuring continuity and compactness. The structure of eGRFC-InFS evolves through participatory learning and a procrastination approach, allowing the addition, merging, deletion, and updating of rules as new data arrive. We evaluate eGRFC-InFS on nine benchmark datasets with varying dimensionality, number of instances, classes, and class proportions. Its performance is compared with those of six alternative evolving classifiers using pairwise T-tests. Results show that eGRFC-InFS outperforms or matches alternative methods, achieving an accuracy improvement of at least 6% on the studied datasets. It has proven to be a reliable and effective solution for classifying data streams.