Back to Deep Learning Mind Map
中文·English
🧠 Deep LearningID: gcn

Graph Conv Network

GCN 谱域
🎯Core Definition
GCN (Graph Convolutional Network, Kipf & Welling 2017) propagates as H(l+1)=σ(D~1/2A~D~1/2H(l)W(l))H^{(l+1)} = \sigma(\tilde{D}^{-1/2}\tilde{A}\tilde{D}^{-1/2}H^{(l)}W^{(l)}), where A~=A+I\tilde{A} = A + I adds self-loops, D~\tilde{D} is its degree matrix, W(l)W^{(l)} is learnable, and σ\sigma is an activation — essentially “degree-normalized weighted average of neighbor features + linear transform”. Spectral motivation: the graph convolution xgθ=Ugθ(Λ)UTxx * g_\theta = U g_\theta(\Lambda) U^T x requires O(n2)O(n^2) eigen-decomposition; approximating gθ(Λ)k=0KθkTk(Λ^)g_\theta(\Lambda) \approx \sum_{k=0}^{K} \theta_k T_k(\hat{\Lambda}) with Chebyshev polynomials and truncating at first order (K=1K = 1), sharing the two parameters (θ0=θ1=θ\theta_0 = -\theta_1 = \theta) and renormalizing (A~=A+I\tilde{A} = A + I) yields exactly the formula above — each node aggregates its 1-hop neighbors, and stacking LL layers expands the receptive field to LL hops. The cost is oversmoothing: after many layers node representations collapse toward the dominant eigenvectors and become indistinguishable (repeated low-pass filtering erases high-frequency information), so classic GCNs are typically 2–3 layers; mitigations include residual connections, JK-Net (jump connections with pooling), PairNorm, and layer normalization.
💡Use Cases
semi-supervised node classification (Cora/Citeseer/PubMed benchmarks), representation learning on citation and social networks; interview staples: writing the propagation formula, the first-order Chebyshev spectral derivation, normalization choices, and the oversmoothing mechanism.
Key Problems Solved
naive graph convolution requires the O(n2)O(n^2) eigen-decomposition of a dense UU that is also non-transferable. The first-order Chebyshev truncation reduces the filter to a single sparse matrix multiplication D~1/2A~D~1/2\tilde{D}^{-1/2}\tilde{A}\tilde{D}^{-1/2} (O(E)O(|E|)), leaving only WW to learn like an ordinary DNN; symmetric normalization stabilizes numerics and enables generalization to unseen nodes, laying the “message passing” paradigm that GAT/GraphSAGE/GIN later follow.
🎯5 High-Frequency Exam Points
1
Write the GCN propagation H(l+1)=σ(D~1/2A~D~1/2H(l)W(l))H^{(l+1)} = \sigma(\tilde{D}^{-1/2}\tilde{A}\tilde{D}^{-1/2}H^{(l)}W^{(l)}) and explain the roles of A~=A+I\tilde{A} = A + I, the symmetric normalization, and W(l)W^{(l)}.
2
Spectral derivation: why is the full graph convolution infeasible (O(n2)O(n^2) eigen-decomposition, dense UU)? How does the first-order Chebyshev approximation gθθ0I+θ1D~1/2A~D~1/2g_\theta \approx \theta_0 I + \theta_1 \tilde{D}^{-1/2}\tilde{A}\tilde{D}^{-1/2} reduce step by step to the GCN formula?
3
Why symmetric normalization D~1/2A~D~1/2\tilde{D}^{-1/2}\tilde{A}\tilde{D}^{-1/2} instead of the random-walk normalization D~1A~\tilde{D}^{-1}\tilde{A}? Differences in numerical stability and aggregation semantics (weighted average vs normalized sum).
4
Mechanism of oversmoothing: why do node representations converge as layers deepen (repeated low-pass filtering, dominant spectral components)? What are the mitigations?
5
Transductive vs inductive learning: why is classic GCN transductive? How does it generalize to unseen nodes/graphs at inference?
Updated 2026-08-12
🎯
Test Your Knowledge: Practice Questions for "Graph Conv Network"
Single choice pitfall questions with instant feedback and mistake tracking.
🚀 Start Card Practice
Previous CardGraph RepresentationNext CardGraph Attention Network

🔗 More Deep Learning Knowledge Cards

Activation FunctionsAdam & AdamWAutograd Compute GraphBatch Normalization