Signals in real world are usually corrupted by noise caused by device malfunction or transmission loss, which needs to be removed before further processing and analysis. Traditional signal denoising methods always adopted smoothing within sliding windows which required a long period of signals as input, while the smoothness also eliminated the variation in signals themselves. Recently, more and more deep learning methods were applied in the denoising task. These methods are almost all supervised where clean data were needed for model training, but we can hardly get the absolutely clean ones in real applications. In contrast, we propose a new unsupervised deep learning framework, \(\infty \) -Net, to tackle the time-series denoising problem on graphs. Specifically, the graph blind-spot network is proposed to incorporate the temporal and spatial correlation for setting the clean signals apart from noise without making any assumption on the distribution of noise. By only using adjacent two frames each time, our model can be nearly online. Through experiments, our method is proved to achieve better performance than that of all peer methods compared for the online graph time-series denoising.