Entity resolution (ER) is a key aspect of data quality and is necessary for big data processing. Existing ER research focuses on data object similarity algorithms, blocking and supervised ER technologies, but pays little attention to matching decision problems in unsupervised ER. This paper proposes a clustering algorithm for ER to complement existing work. The algorithm builds a weighted similarity graph with data objects and their pairwise similarities. During clustering, the similarity between a cluster and a vertex is dynamically computed via random walk with restarts on the similarity graph. The basic logic behind clustering is that a cluster absorbs the nearest neighbor vertex iteratively. A data object ordering method is also proposed to optimize clustering order, promoting clustering accuracy. Further, an improved computation method of random walk's stationary probability distribution is proposed to reduce cost of the clustering algorithm. The evaluation on real datasets and synthetic datasets validates effectiveness of the proposed algorithm.