Hybrid Deep Sequential Learning and IBSCA-Driven Feature Selection for Robust Classification of Imbalanced Dataset

Abstract

In many real-world datasets, class imbalance is still an issue that skews traditional classifiers in favour of majority classes and hinders the performance of minority classes. This work introduces a novel end-to-end pipeline that combines IBSA feature selection (92% dimensionality reduction), BiGRU-LSTM-Attention hybrid modelling for complex spatiotemporal pattern capture, and adaptive thresholding for fairness optimisation. With state-of-the-art results of 98.61% accuracy/99.92% sensitivity and 99.79% accuracy/99.89% sensitivity, the proposed BiGRU-LSTM model outperforms optimization-enhanced hybrids by 1-2% and SMOTE baselines by 7-15% on benchmark imbalanced datasets covering binary and multi-class scenarios. These notable advancements show how well the architecture handles severe class imbalance across domains, establishing new performance standards that could be applied to federated learning and high-dimensional data processing.