Graph Consistency Regularization (GCR) is a novel framework that promotes class-aware and meaningful feature representations by injecting a relational graph structure derived from model predictions into the learning process. GCR acts as a kind of self-prompting mechanism, using its own output to improve its internal structure. GCR introduces parameter-free Graph Consistency Layers (GCLs) at arbitrary depths to address noisy inter-class similarities that contradict the model's prediction semantics. Each GCL builds a batch-level feature similarity graph and aligns it with a global class-aware mask prediction graph derived by adjusting the softmax prediction similarity as a within-class metric. This alignment acts as a semantic regularizer across the network, forcing feature-level relationships to reflect class-consistent prediction behavior. GCR introduces a multi-layer, cross-spatial graph alignment mechanism and uses adaptive weights to learn layer importance from graph disparity magnitude. GCR improves feature quality by prioritizing semantically reliable layers and suppressing noisy layers without modifying the model architecture or training procedure. GCR is model-agnostic, lightweight, and improves semantic structure across diverse networks and datasets.
