© 2022 IEEE.We study the problem of estimating time-varying graph signals from missing observations. We propose a method based on learning graph dictionaries specified by a set of time-vertex kernels in the joint spectral domain. The parameters of the time-vertex kernels are optimized jointly with the sparse representation coefficients of the signals, so that the learnt representation fits well to the available observations of the time-vertex signals at hand. The missing observations of the signals are then estimated based on their reconstruction with the learnt model. Experimental results on real graph signal data sets show that the proposed method outperforms classical graph-based regression approaches.