Abstract: Aiming at  the problems that the traditional K-Means algorithm was sensitive to initial centroids, prone to local optima, and failing to fully mine the potential semantic features of clustering results, we proposed a  K-Means clustering algorithm based on heuristic crossover strategy optimization. Firstly, the algorithm used  a density-driven heuristic crossover initialization strategy to screen representative parent points in high-density regions, and  introduced a crossover coefficient to dynamically generate diverse initial centroids to  reduce the volatility of clustering results caused by random initialization. Secondly, during the clustering iteration process, by combining the information of parent points with the intra-cluster mean update rule, the centroid positions were dynamically adjusted through crossover operations, which solved the problem of inter-cluster overlap caused by the local optima of the traditional algorithm. Finally, the optimized clustering results were input into a multi-layer perceptron, which utilized its nonlinear mapping ability to mine potential features and  achieved  deep fusion of clustering results with  deep semantic features. Experimental results show that the contour coefficient, Davies-Bouldin index, and adjusted Rand index of the algorithm reach 0.634, 1.398 and 0.621, respectively, which are significantly superior to other improved algorithms, effectively improving clustering accuracy, stability, and interpretability of the algorithm.

Key words: heuristic crossover strategy, K-Means clustering algorithm, multi-layer perceptron, feature fusion