Traditional clustering algorithms are sensitive to initial clustering centers, and their clustering effect is sometimes poor. For these reasons, we present a Gk-prototypes clustering algorithm to process numerical properties and classification properties. Based on the classical k-prototypes clustering algorithm, the proposed algorithm uses the de-fuzzy similarity matrix to construct coarse particle sets, and employs particle calculation and the maximum and minimum distance method to determine the initial clustering center, thus the objective function is improved. Experimental results and theoretical analysis show that the Gk-prototypes clustering algorithm is more accurate, more effective and more robust than other improved algorithms based on k-prototypes.