Graph neural networks (GNNs) classify a node by repeatedly averaging its features with those of its neighbors. This works when linked nodes tend to share a label, a property called homophily. It fails in two related ways. On heterophilous graphs, where neighbors often belong to other classes, averaging washes out the evidence that distinguishes a node. In deep networks, repeated averaging drives all node representations toward a nearly identical, low-dimensional state, which is over-smoothing. Most fixes change how neighbors are weighted, using signs, high-pass filters, or learned edge maps. But the signal that decides the weighting is itself computed from representations that are being smoothed, so the evidence about whom to trust degrades exactly as the network gets deeper.
We propose computing that evidence once, before any propagation, as a separate per-node view that depth cannot erode. For each node we build its two-hop neighborhood graph (its "ego-graph"). We add edges in order of feature dissimilarity and record how groups of neighbors merge and when loops form. The standard tool for recording this is persistent homology, which tracks connected components and cycles as a threshold sweeps. The resulting summary tells a node more than "my neighbors look unlike me." It also tells it whether those unlike neighbors form one coherent group, which is usable, consistent evidence, or a scatter of unrelated nodes, which is noise. We use this summary to set, per node and per layer, how much the network mixes neighbor messages into the node's own representation. The claim to test is that this neighbor-to-neighbor structure predicts which nodes to trust better than center-to-neighbor similarity does. If it does, it should keep deep networks accurate and high-rank on heterophily benchmarks that are free of data leakage.
Core Claim
Existing heterophily-aware models judge each neighbor pairwise against the center node. GGCN learns signed edge weights from cosine similarity between a node and each neighbor (Two Sides of the Same Coin). Filterbank GNNs learn global scalars that balance low-pass averaging against high-pass identity terms (Complete the Missing Half). Neural sheaf diffusion learns a linear map on each edge (Neural Sheaf Diffusion). None of these asks how a node's neighbors relate to one another. Yet that relation separates two situations that look the same pairwise. In the first, a node is surrounded by one tight group of another class, as in bipartite-like fraud or word-adjacency structure; its messages carry a stable, usable signal. In the second, its neighbors are mutually unrelated, and their messages add noise. The 0-dimensional persistence of a feature-filtered ego-graph encodes exactly this difference.
Our claim is that a gate driven by this precomputed, non-propagated topological view improves two things over the strongest simple model on clean heterophily benchmarks. It raises accuracy, and it keeps representations high-rank as depth grows. The improvement should survive replacing the topology with center-to-neighbor similarity statistics of equal size, and it should vanish when the topology is shuffled across nodes.
This also answers a gap the literature leaves open. Topological multi-view models such as TopoGCL (TopoGCL) and TensorMV-GCL (Tensor-Fused Multi-view Graph Contrastive Learning) fuse persistence summaries with a GCN channel, but only for whole-graph classification on small molecular datasets. We move the topological view to the node level and give it a job: deciding propagation, rather than serving as one more embedding to concatenate.
The Topological View
For node $v$ with two-hop ego-graph $G_v$, give each edge $(a,b)$ the weight $w_{ab} = 1 - \cos(x_a, x_b)$, where $x$ are raw input features. This makes edge weights measure dissimilarity. The filtration adds edges in increasing $w_{ab}$. We also run a second filtration that gives each vertex $u$ the value $1-\cos(x_u,x_v)$, its dissimilarity to the center. We record 0-dimensional persistence, meaning when neighbor groups merge, and 1-dimensional persistence on the graph's cycles. The resulting diagrams are vectorized with persistence landscapes, fixed-length functional summaries of the kind TopoGCL uses, into a vector $t_v$ of about 32 dimensions.
Computing the 0-dimensional part needs only sorting plus union-find over the ego-graph's edges. It therefore sidesteps the cost of building clique complexes that the expressivity study flags as prohibitive (On the Expressivity of Persistent Homology in Graph Learning). Ego-graphs of very high-degree nodes are subsampled to a fixed edge budget. The filtration is fixed and cheap, which is a deliberate choice: Going beyond persistent homology using persistent homology shows that vertex-based and edge-based filtrations capture different information. That is why we use one of each rather than trying to learn a filtration.
The Model
The backbone is the model that A critical look at the evaluation of GNNs under heterophily found hardest to beat. It is a residual GNN with ego-neighbor separation, meaning the node's own representation and its aggregated neighbors get separate weight matrices. We change one thing. A small network reads $t_v$, which is fixed, and outputs per-node, per-layer coefficients:
$$h_v^{(l+1)} = h_v^{(l)} + \sigma\Big(W_1^{(l)} h_v^{(l)} + g_l(t_v)\, W_2^{(l)} \textstyle\sum_{u \in N(v)} \tilde a_{vu} h_u^{(l)}\Big),$$
Here $h_v^{(l)}$ is node $v$'s representation at layer $l$, $W_1^{(l)}, W_2^{(l)}$ are the separated ego and neighbor weights, $\tilde a_{vu}$ are normalized adjacency weights, $\sigma$ is a nonlinearity, and $g_l(t_v) \in [-1, 1]$ is the gate. The gate can switch neighbor messages off or flip their sign for nodes whose ego-topology signals noise. Because $t_v$ is never propagated, the gate cannot suffer the "pseudo-heterophily" drift that Two Sides of the Same Coin identifies. That drift is a node's representation crossing a class boundary in deep layers and then misleading its neighbors, which in turn corrupts similarity-based weights computed from $h$.
Research Program
The work has three parts. The first is a mechanism study on a synthetic generator. We plant graphs where every node has the same distribution of center-to-neighbor similarities, but neighbor sets are either coherent groups of one foreign class or scattered mixtures, and the label depends on which. By construction, center-to-neighbor statistics carry no signal there, while ego-graph persistence does. If the gated model cannot exploit this, the idea is wrong at its core.
The second part is a depth study on clean benchmarks at depths 2, 4, 8, 16, 32, and 64. At every layer we record the effective rank of the representation matrix, the exponential of the entropy of its normalized singular values. Are we measuring oversmoothing in graph neural networks correctly? reports an average Pearson correlation of 0.91 between effective rank and accuracy, while Dirichlet energy, the usual smoothness measure, showed near-zero or negative correlation in trained models. We therefore report effective rank, never energy alone, as the over-smoothing measure. The same paper warns that high rank does not guarantee usefulness, since pure noise has full rank, so rank is always read alongside accuracy.
The third part is analysis. We stratify accuracy by node-level homophily and by the coherence score that $t_v$ encodes. We also inspect what the gates learned, to see whether they close on incoherent neighborhoods as predicted.
Baselines
The naive reference is a graph-agnostic ResNet on node features. Platonov et al. found that many heterophily-specific models fail to beat it on clean data (A critical look at the evaluation of GNNs under heterophily). Any gain from structure must first clear this bar.
The strongest existing competitor is the ego-neighbor-separated GNN (GraphSAGE-style and GAT-sep) from the same paper. It took 13 of the top 15 places on the new suite. The file reports no per-dataset numbers for it: no reported result found on Lacuna for roman-empire, amazon-ratings, minesweeper, tolokers, or questions. We will therefore rerun it ourselves with Platonov et al.'s splits and metrics: accuracy for roman-empire and amazon-ratings, and ROC-AUC for minesweeper, tolokers, and questions. It is the backbone of our model, which makes the comparison a clean one-component ablation. Beating it means a mean gain larger than the standard deviation across Platonov's 10 splits, on at least three of five datasets, at matched parameter count, with no loss at depth 2.
Among specialized models we include GGCN, O(d)-NSD, filterbank GCN, and the multi-scale Haar model HMH (Hierarchical Multi Scale Graph Neural Networks). GGCN is the most direct competitor because it also uses similarity to sign messages. NSD and HMH are the published models claiming both heterophily robustness and stability to 64 layers. The multi-track model MTGCN (Multi-Track Message Passing) is included where its memory, which grows with the number of classes, allows. The best numbers the file reports for these models come from the leaky legacy suite: MTGCN at 92.10% accuracy on Texas and 90.20% on Wisconsin, GSAN at 61.2% on Chameleon (Geometric Scattering Attention Networks), and GGCN with an average rank of 1.78 across nine datasets. We cite these only for context, because Platonov et al. showed that chameleon and squirrel contain more than 50% duplicated nodes. All baselines receive the same hyperparameter search budget (identical trial counts over learning rate, dropout, width, and depth), the same splits, and the same hardware. The topology precomputation time is reported and counted.
Two ablations isolate the key ingredient. The first, a similarity-histogram gate, replaces $t_v$ with a histogram of equal dimension built from center-to-neighbor cosine similarities. This is the GGCN-style information without neighbor-to-neighbor structure. The second, a shuffled-topology gate, permutes $t_v$ across nodes. A third variant feeds $t_v$ only as an extra input feature with no gating, to test whether the gating role matters or the view is just a useful descriptor.
Why not just…?
Why not just use ego-neighbor separation, which already wins on clean benchmarks? This is the real threat, and the design is built around it. Separation keeps a node's identity from being washed out, but it applies the same neighbor-mixing weights to every node. It cannot tell a coherent foreign-class neighborhood from a noisy one. Our model is that exact baseline plus a gate, so the experiment answers the question directly. If the gate gives no gain on the clean suite and the synthetic generator, we conclude that separation suffices and report it.
Why not just gate on center-to-neighbor similarity, as GGCN does? Such a gate is cheaper and may capture most of the benefit. The similarity-histogram ablation tests exactly this at equal dimension. The synthetic generator, where center-to-neighbor statistics are uninformative by construction, shows whether neighbor-to-neighbor structure adds information the histogram cannot. If the two tie on real data, the honest conclusion is that real heterophilous graphs rarely contain the coherent-versus-scattered distinction. That finding is still worth reporting.
Why not just drop the graph and use an MLP or ResNet? On some clean datasets this may genuinely be competitive, which is why it is the first reference point. Our stratified analysis asks where structure helps at all. We expect gains to concentrate on roman-empire and minesweeper, where labels depend on local arrangement, and to be small on amazon-ratings.
Why not just concatenate multi-hop features, as MixHop does? Multi-hop features show a node what lies two steps away. They do not tell it whether to trust what it sees, and powers of the adjacency matrix still mix classes indiscriminately, as Multi-Track Message Passing argues. We include a two-hop concatenation baseline on the separated backbone at matched width.
Decisive Checks
We evaluate on the five datasets of Platonov et al., the deduplicated chameleon and squirrel they provide, and Cora, Citeseer, and Pubmed as homophilic controls, where the gate should learn to stay open and cost nothing. The idea is supported only if all of the following hold. The gated model beats the separated backbone and the similarity-histogram gate on the clean suite under matched budgets. The shuffled-topology control erases the gain. Effective rank at depth 32–64 stays measurably higher than the backbone's, and does so together with higher accuracy. The synthetic generator shows a clear gap between topological and histogram gates. If only the synthetic check passes, the mechanism is real but rare in practice, and we will say so. If the depth curves show that the separated backbone already avoids rank collapse on clean data, that negative result is itself useful. It would mean the 64-layer anti-over-smoothing claims made on the legacy suite solved a problem that clean benchmarks do not pose.
References
- Two Sides of the Same Coin: Heterophily and Oversmoothing in Graph Convolutional Neural Networks
- Complete the Missing Half: Augmenting Aggregation Filtering with Diversification for Graph Convolutional Networks
- Neural Sheaf Diffusion: A Topological Perspective on Heterophily and Oversmoothing in GNNs
- TopoGCL: Topological Graph Contrastive Learning
- Tensor-Fused Multi-view Graph Contrastive Learning
- On the Expressivity of Persistent Homology in Graph Learning
- Going beyond persistent homology using persistent homology
- A critical look at the evaluation of GNNs under heterophily: are we really making progress?
- Are we measuring oversmoothing in graph neural networks correctly?
- Hierarchical Multi Scale Graph Neural Networks: Scalable Heterophilous Learning with Oversmoothing and Oversquashing Mitigation
- Multi-Track Message Passing: Tackling Oversmoothing and Oversquashing in Graph Learning via Preventing Heterophily Mixing
- Geometric Scattering Attention Networks