Generative model controls heterophily in graph signals.
problem Controlling heterophily in graph signals for better model effectiveness.
method Combines graphon-based generator with spectral filtering of Gaussian node features.
result Establishes theoretical guarantees for heterophily control and convergence.
A new GNN framework for graphs with heterophily.
problem Graphs with heterophily (nodes from different classes often connected).
method CPGNN framework with an interpretable compatibility matrix.
result CPGNN achieves state-of-the-art results in heterophily settings.
The paper explores how different patterns of heterophily affect Graph Neural Networks.
problem Understanding the impact of heterophily on Graph Neural Networks.
method Theoretical analysis and experiments with Heterophilous Stochastic Block Models (HSBM).
result The impact of heterophily on classification depends on the Euclidean distance of neighborhood distributions and the averaged node degree.
Graph neural networks struggle with heterophily, but new designs improve their performance.
problem Graph neural networks struggle with heterophily (networks where connected nodes may have different class labels and dissimilar features).
method Ego- and neighbor-embedding separation, higher-order neighborhoods, and combination of intermediate representations.
result The identified designs increase the accuracy of GNNs by up to 40% and 27% over models without them on synthetic and real networks with heterophily, respectively.
New method handles structural uncertainty in graphs better than existing models.
problem Handling heterophily and structural noise in semi-supervised learning on graphs.
method Sparse signed message passing network that models a posterior distribution over signed adjacency matrices.
result Our method outperforms strong baseline models on heterophilic benchmarks under both synthetic and real-world structural noise.
Heterophily affects GNN robustness; separating ego- and neighbor-embeddings improves defense.
problem The robustness of GNNs to adversarial attacks.
method Formalized relation between heterophily and GNN robustness; empirical analysis; design principles for improved robustness.
result Separating ego- and neighbor-embeddings increases GNN robustness.
PGMs and GNNs differ in capturing network data; PGMs outperform GNNs in noisy and heterophily scenarios.
problem Comparing PGMs and GNNs in network data.
method Link prediction task with synthetic and real networks; three experiments on input features, noise, and heterophily.
result PGMs outperform GNNs in noisy and heterophily scenarios.
MPNNs struggle with class-bottlenecks and heterophily, leading to performance limitations.
problem Performance limitations of MPNNs under heterophily and structural bottlenecks.
method A statistical framework decomposing model performance into SNR components and proving bounds on sensitivity.
result Optimal graph structures for maximizing higher-order homophily are disjoint unions of single-class and two-class-bipartite clusters.
Sheaf Neural Networks improve graph learning with geometric insights.
problem Graph heterophily and over-smoothing issues.
method Inspired by Riemannian geometry, computes sheaves using orthogonal maps.
result Achieves promising results with reduced computational overhead.
PolyNSD improves Neural Sheaf Diffusion with polynomial operators and spectral rescaling.
problem Limitations of common Neural Sheaf Diffusion implementations, including scalability and stability issues.
method Introduces Polynomial Neural Sheaf Diffusion (PolyNSD) with a degree-K polynomial propagation operator and spectral rescaling.
result PolyNSD achieves state-of-the-art results on both homophilic and heterophilic benchmarks with reduced runtime and memory requirements.
Adaptive GPR-GNN optimizes node feature and topology learning.
problem Optimizing GNNs for both node features and graph topology, regardless of homophily or heterophily.
method Adaptive Universal Generalized PageRank (GPR) Graph Neural Network (GPR-GNN) that learns optimal GPR weights.
result Significant performance improvement on node classification tasks compared to state-of-the-art GNNs.
Paper proposes JDR to denoise graph features and rewire graphs for better node classification.
problem Jointly denoise noisy graph features and rewire graphs for improved node classification.
method Align leading spectral spaces of graph and feature matrices to solve non-convex optimization problem.
result JDR consistently outperforms existing methods on various node classification tasks.
GCNs learn by embedding similar nodes within a class and leveraging consistent neighborhood structures.
problem Understanding how GCNs perform semi-supervised node classification on both homophilous and heterophilous graphs.
method Investigated the latent node embeddings and neighborhood structures of GCNs.
result GCNs learn by embedding similar nodes within a class and leveraging consistent neighborhood structures.
Graphs models are vulnerable to distribution shifts, which this work explains and mitigates.
problem Graph Neural Networks (GNNs) are susceptible to distribution shift, leading to performance degradation.
method Theoretical analysis quantifying conditional shift, proposing an approach to estimate and minimize it.
result The proposed approach demonstrates up to 10% absolute ROC AUC improvement under various distribution shifts.
GCNIII combines Wide & Deep for better node classification.
problem Issues with graph convolutional networks in node classification tasks.
method Proposes GCNIII framework integrating Wide & Deep architecture and three techniques.
result Demonstrates improved performance in various node classification tasks.
SCNode improves node embeddings for GNNs in both homophilic and heterophilic graphs.
problem Challenges in node representation quality and generalization in GNNs, especially in heterophilic graphs.
method SCNode integrates spatial and contextual information to create more discriminative and structurally aware node embeddings.
result SCNode achieves superior performance over conventional GNN models on benchmark datasets.
GNNs generalize better on homophilic graphs than heterophilic ones.
problem Understanding the generalization error of GNNs on graph data.
method Analytical tools from statistical physics and random matrix theory.
result Risk is shaped by graph noise, feature noise, and training labels.
GCNs perform well on heterophilous graphs under certain conditions.
problem The necessity of homophily for good GNN performance.
method Empirical evaluation and theoretical analysis of GCNs on heterophilous graphs.
result GCNs can achieve strong performance on heterophilous graphs under certain conditions.
Causal inference improves heterophilic graph learning.
problem Capturing asymmetric node dependencies in graph learning.
method Intervention-based causal inference for graph structure learning.
result CausalMP achieves superior link prediction performance.
FSD-CAP improves graph feature imputation under high missing rates.
problem Challenges in imputing missing node features in graphs, especially under high missing rates.
method Two-stage framework: subgraph expansion, fractional diffusion, class-aware propagation.
result Significantly improved imputation quality compared to existing methods, achieving high accuracy on benchmark datasets.
GNNGuard defends Graph Neural Networks against structural perturbations.
problem Adversarial attacks on graph neural networks can degrade performance catastrophically.
method Detects and quantifies the relationship between graph structure and node features, then uses this to mitigate attacks.
result GNNGuard outperforms existing defenses by 15.3% on average across various attacks and datasets.
Improved graph neural networks by separating feature aggregation and depth.
problem Understanding feature importance in graph neural networks without prior information.
method Decoupling feature aggregation and depth, using softmax as a regularizer, and introducing 'Soft-Selector' and 'Hop-Normalization'.
result FSGNN model achieves up to 64% accuracy improvements in node classification tasks.
Simplifies GNN models by selecting important features for node classification.
problem Challenges in analyzing and selecting important features in GNN models.
method Decoupling feature aggregation and depth, using softmax and L2-normalization.
result FSGNN achieves comparable or higher accuracy than state-of-the-art GNN models.