Exploiting label relationship to help improve learning performance is important for multi-label learning. The classifier chain method and its variants are shown to be a powerful solution to such a problem. However, its learning process requires the ordering of labels, which is hard to obtain in real-world situations, and incorrect label ordering may cause a suboptimal performance. To overcome the drawback, this paper presents a classifier circle method for multi-label learning. It initializes the label ordering randomly, and then subsequently and iteratively updates the classifier for each label by connecting the labels as a circle. Experimental results on a number of data sets show that the proposal outperforms classifier chains method as well as many state-of-the-art multi-label methods.