Randomly allocating initial centroids may lead to undesired steady states for fuzzy c-means (FCM) clustering. This paper proposes an alternative method to automatically search initial centroid location based on data density. Specifically, this method auto-searches points located in high-density domains as centroids using directed acycline graph (DAG) based algorithm, and then iteratively fnding the optimal patterns. Compared with random initialization method, our method seems to have the potential to improve FCM accuracy for larger data size with seconds' tradeoff in computational time using published datasets.