This work tries to establish the connection between Bayesian inference over a graphical model and Graph neural networks. With such a connection, we are able to measure the values of nonlinear operations in GNNs for the node classification task. We obtain a “negative” result: When node attributes are not very informative, non-linear operations during the message are kind of useless, which matches many previous empirically successful architectures such as APPNP, GPRGNN and spectral GNNs such as JacobiConv, etc.