We propose graph consistency regularization, a method that improves classification by aligning feature and prediction structures to suppress inter-class noise and strengthen intra-class cohesion.