Clustering is one of the focused problems in multimedia data mining, and similarity measurement among data is fundamental to clustering. In multimedia data clustering, the corresponding vector features are always of high dimensionality. Most traditional measurement methods, however, are only efficient for low dimensional data. This paper, based on an analysis of general characteristics of data presented in high dimensional spaces, proposes a new similarity measurement for multimedia data mining. It used a special strategy to split the original data space before computing the similarity among data points, thus efficiently avoiding the influence of noisy data in high dimensional dimensional spaces. Experiments show that the new method presented is effective.