Sparse multi-objective optimization problems (SMOPs) frequently exist in a variety of disciplines such as machine learning, economy, and signal processing. Evolutionary algorithms have demonstrated their proficiency in optimizing complex problems in recent years, although their performance often deteriorates significantly on large-scale SMOPs. In an effort to accelerate the convergence, this paper suggests a multi-granularity variable clustering method for evolutionary algorithms. This method estimates the sparse distribution of decision variables at each generation and partitions them into a varying number of layers, each with a distinct probability of being zero. These clustering outcomes inspire the development of a crossover operator and a mutation operator, which prove adept at efficiently generating sparse solutions. Experimental evaluations on both benchmark and real-world SMOPs confirm that an evolutionary algorithm incorporating the new crossover operator and mutation operator converges more rapidly than its state-of-the-art counterparts.