Abstract
To address the instability and performance issues of the classical K-Means algorithm when dealing with massive datasets, we propose SOSK-Means, an improved K-Means algorithm based on Spark optimization. SOSK-Means incorporates several key modifications to enhance the clustering process.Firstly, a weighted jump-bank approach is introduced to enable efficient random sampling and preclustering. By incorporating weights and jump pointers, this approach improves the quality of initial centers and reduces sensitivity to their selection. Secondly, we utilize a weighted max-min distance with variance to calculate distances, considering both weight and variance information. This enables SOSK-Means to identify clusters that are farther apart and denser, enhancing clustering accuracy. The selection of the best initial centers is performed using the mean square error criterion. This ensures that the initial centers better represent the distribution and structure of the dataset, leading to improved clustering performance. During the iteration process, a novel distance comparison method is employed to reduce computation time, optimizing the overall efficiency of the algorithm. Additionally, SOSK-Means incorporates a Directed Acyclic Graph (DAG) to optimize performance through distributed strategies, leveraging the capabilities of the Spark framework. Experimental results show that SOSK-Means significantly improves computational speed while maintaining high computational accuracy.