Feature selection has been explored extensively for several real-world applications. In this paper, we address a new solution of selecting a subset of original features for unlabeled data. The concept of our feature selection method is referred to a basic characteristic of clustering in thata data instance usually belongs in the same cluster with its geometrically nearest neighbors and belongs to different clusters with its geometrically farthest neighbors. In particular, our method uses instance-based learning for quantifying features in the context of the nearest and the farthest neighbors of every instance, such that using salient features can raise this characteristic. Experiments on several datasets demonstrated the effectiveness of our presented feature selection method.