Computational Intelligence Techniques for Pattern Recognition in Complex Data

Abstract

Pattern recognition in complex, high-dimensional, and heterogeneous data remains one of the most challenging and consequential problems in
computational intelligence, with applications spanning genomics, hyperspectral remote sensing, industrial quality control, and anomalous behaviour
detection in cybersecurity. While deep learning has achieved state-of-the-art performance on structured benchmark datasets, it frequently
underperforms on small-sample, high-dimensional, and class-imbalanced complex data regimes where traditional computational intelligence
methods retain comparative advantages. This study presents PatternCI, a unified computational intelligence framework for pattern recognition in
complex data integrating fuzzy rough set-based feature selection, an evolutionary ensemble classifier, and a self-organising map (SOM)-guided
cluster validation module. PatternCI was evaluated across four complex pattern recognition benchmarks: hyperspectral land cover classification
(Indian Pines, 200 spectral bands, 16 classes), genomic cancer subtype identification (TCGA pan-cancer, 20,531 features, 33 cancer types),
industrial surface defect detection (DAGM 2007 dataset, 10 texture classes), and network intrusion detection (CIC-IDS-2017, 78 features, 14 attack
categories). PatternCI achieved classification accuracies of 97.4%, 94.8%, 98.1%, and 99.2% respectively, surpassing deep learning baselines
(CNN, ResNet, MLP) by 2.3–8.7% on small-sample regimes and evolutionary baselines by 4.1–11.3%. Feature selection reduced dimensionality by
73.4% on average while preserving 98.6% of classification information. These results establish PatternCI as an effective framework for computational
intelligence-driven pattern recognition across diverse complex data domains.