Patch Uniform Fusion Transformer with circle-inspired optimized walrus-based feature selection for high-dimensional...

Feature selection approaches have historically been limited by processing performance during high-dimensional data analysis, which presents additional inefficiencies when dealing with bigger datasets. Hybrid search methods are also great and powerful, but they are computationally intensive and do not effectively decrease features, affecting accuracy and processing speed.

To address these issues, this research proposes a unique hybrid feature selection approach called Improved Circle-Inspired Walrus Optimization (ICIWO) for very high-dimensional datasets. The Fully Spiking Variational Autoencoder (FSVA) compresses data using spiking neural networks to minimize dimensionality while retaining important information.

On the ALL-AML dataset, it performs well, with a fitness value of 0.0607, accuracy of 98%, AUC of 0.9913, precision of 99.43%, recall of 99.47%, and F1-score of 99.28%. For the GLI-85 dataset, ICIWO records a fitness value of 0.0727, an accuracy of 97%, an AUC of 0.9857, and a precision of 98.87%, while in the CLL-SUB-111 dataset, it achieves an accuracy of 95% with an AUC of 0.9753.

The results indicate ICIWO's ability to provide a robust, cost-effective solution for high-dimensional data, overcoming the limitations of traditional and existing hybrid methods and enabling improved insights and pattern recognition across various datasets.