The computational complexity of high-dimensional data increases dramatically due to the dimensionality catastrophe, and the traditional feature selection method models face the problems of model overfitting, computational inefficiency, and feature redundancy, especially in big data scenarios such as medical images and text semantics. Existing filtering methods, wrapping methods and embedded methods have the problems of ignoring feature interactions, exponentially increasing computational cost with feature dimensions, and easily falling into local optimization, respectively. Aiming at the above problems, this paper proposes a feature selection algorithm MGCSO for granular spheres based on cockroach swarms and mutual information, which integrates the dynamic perturbation mechanism of CSO, introduces an improved mutual information metric model and an adaptive clustering strategy for granular spheres, and the algorithm realizes effective screening through three-phase synergism. Numerical experiments show that the algorithm exhibits significant advantages in computation time while maintaining good classification accuracy and selecting smaller feature subsets, especially in high-dimensional large datasets.