Political networks carry more than one ideological axis. A legislator, subreddit or Twitter community can be placed left-to-right on economics and, somewhat independently, on a second axis such as libertarian versus authoritarian, or establishment versus populist. Graph neural networks (GNNs) are models that update each node by averaging over its neighbors, and they are increasingly used to infer these positions. This note argues that repeated neighbor averaging does not treat all ideological axes equally. The main partisan axis runs along the network's large communities, so averaging leaves it intact. A cross-cutting axis varies within communities and across bridges, so averaging erases it first. If this holds, part of the familiar "everything collapses to one left-right dimension" problem in computational political science comes from the propagation step, not from the data or the labels.
We propose to measure this "dimension-selective over-smoothing" directly. Over-smoothing is the known tendency of deep GNNs to make all node representations alike. We will test whether a node's local structural role predicts where cross-cutting ideology lives. Structural role here means its position in small subgraph patterns, such as being the middle of a path or the edge of a clique. We will also test whether letting those roles control how much each ideological channel is averaged preserves the second axis without hurting the first. In two sentences: neighbor averaging acts as a filter that keeps the ideology aligned with community structure and deletes the ideology that cuts across it. Nodes in bridging structural positions are where the deleted axis is concentrated, so propagation should be turned down for that axis at those nodes.
Core Claim
Message passing on the normalized adjacency matrix is a low-pass filter. It keeps signals that vary slowly across edges and suppresses signals that change sharply between neighbors. This is standard in the over-smoothing literature. PairNorm: Tackling Oversmoothing in GNNs shows it is a property of the graph convolution operator itself, and Complete the Missing Half adds a high-pass channel to recover what the filter throws away. What nobody has examined is what this filter does to a multi-dimensional target whose axes line up differently with the graph.
For ideological axis $k$ with node scores $x_k$, define its graph roughness as the Rayleigh quotient
$$R_k = \frac{x_k^\top \Delta_{norm}\, x_k}{x_k^\top x_k},$$
where $\Delta_{norm} = I - D^{-1/2} A D^{-1/2}$ is the normalized Laplacian and $x_k$ is degree-rescaled to match that operator. $R_k$ is small when neighbors agree on axis $k$ and large when they disagree. Propagation multiplies each spectral component with Laplacian eigenvalue $\lambda$ by $(1-\lambda)$ per layer. To first order, then, the fraction of axis $k$ that survives $L$ layers of weight-free propagation shrinks roughly like $(1-R_k)^{L}$. The prediction is falsifiable: axes with larger roughness are lost faster with depth, and on polarized networks the partisan axis has small $R_k$ while cross-cutting axes have large $R_k$.
The second claim concerns subgraph patterns. Cross-cutting positions should concentrate at nodes in bridging roles. Surrogate explanations for role discovery on graphs found that boundary-spanning papers in a citation graph sit disproportionately in path-center and clique-periphery orbits. An orbit is a node's symmetry position inside a small subgraph pattern, such as the middle versus the end of a three-node path. If the same holds for ideology, orbit counts tell us where to stop averaging on the cross-cutting channel.
Research Program
The program has three parts, and the first carries most of the weight.
The first part is a measurement study. We take graphs where a two-or-more-dimensional ideological ground truth exists, compute $R_k$ per axis, and track how well each axis can be linearly decoded from GNN representations as depth grows. Measurement follows Analysis of Dirichlet Energies as Over-smoothing Measures. That paper shows that pairing a normalized propagation operator with an unnormalized energy metric produces artifacts on non-regular graphs, and political graphs are strongly non-regular. It also shows that representations collapsing in direction (over-smoothing) must be separated from representations shrinking in norm (over-shrinking). We therefore report energy under the Laplacian matching the operator, alongside the Frobenius norm. Our per-axis Dirichlet energy is a new diagnostic. Existing over-smoothing measures summarize the whole representation, and Are we measuring oversmoothing in graph neural networks correctly? shows how fragile such summaries are.
The second part is a model we call role-gated facet propagation. Each node carries $K$ output channels, one per ideological axis, supervised directly against that axis. At every layer, channel $k$ of node $i$ mixes its own previous state with its neighborhood average. The mixing weight is a sigmoid of a learned linear function of the node's orbit-count vector: log-scaled counts of the 2- to 4-node graphlet orbits, which stay cheap to compute on sparse graphs. Each axis thus learns its own map from structural role to "how much should this node listen to its neighbors." On the partisan channel, the weights should favor full averaging. On the cross-cutting channel, bridge-role nodes should keep their own evidence.
The model borrows orbit counting from Improving Graph Neural Network Expressivity via Subgraph Isomorphism Counting, but uses it differently. There, substructure counts are added as extra input features to break symmetries in shallow models. Here they control the propagation strength of each channel. It also differs from Independence Promoted Graph Disentangled Networks (IPGDN), which splits aggregation into independent latent channels by feature-based routing with no link to named axes. Our channels are tied to measured ideological dimensions, and routing depends on structure. Finally, it differs from When Do GNNs Work, which gates aggregation on feature similarity and neighborhood entropy. That gate is shared across all output dimensions, which by our hypothesis is exactly the wrong granularity.
The third part is interpretation. We read the learned gate weights per axis and test whether cross-cutting channels load on the same bridge orbits that SERD identified. We also test whether ideological cross-pressure in parliament shows the same structural signature. Discovering Opinion Intervals from Conflicts in Signed Graphs found overlapping positions for parties such as the FDP, which is the kind of case we expect to appear.
The application setting constrains the design. Several political GNNs (Legislator Representation Learning with Social Context and Expert Knowledge; Modeling Ideological Agenda Setting and Framing in Polarized Online Groups) add explicit homophily or echo-chamber losses that pull connected nodes together. By our argument, such losses speed up the erasure of cross-cutting axes, so we treat them as a condition to test rather than a default. Those models are also shallow. We therefore do not assume political GNNs must be deep. Our claim is that the damage is axis-specific and appears within the 2–4 layers these systems actually use, and the measurement study checks this.
Baselines
The naive reference points are two. The first is a text-only or feature-only predictor with no graph, a linear probe on node text features, which sets the floor any graph method must beat on each axis. The second is a plain GCN at matched depth. For the text-only reference on the first axis, the strongest reported number in our file comes from Linear Representations of Political Perspective Emerge in Large Language Models. Linear probes on attention heads predict first-dimension DW-NOMINATE scores for 116th-Congress lawmakers with Spearman correlation of about 0.85 for the best single head and 0.87–0.88 for an ensemble of 32 heads. We found no reported result on Lacuna for the second DW-NOMINATE dimension from text probes or from GNNs. Establishing that number is part of the contribution. We will run the same probe style on the same lawmakers so the graph methods can be compared against a strong text-only reference on both axes.
The strongest existing graph competitor on the legislator task is the Gated R-GCN of Legislator Representation Learning with Social Context and Expert Knowledge. It reports 91.30% accuracy on SemEval-2019 perspective detection, a different task, and no extracted number for roll-call prediction. We re-run it with its relational structure and text initialization, trained on our per-axis regression targets, with and without its echo-chamber loss. Two over-smoothing remedies act as generic competitors. PairNorm keeps total pairwise distance between node representations fixed. The diversification filterbank of Complete the Missing Half runs a high-pass channel next to the usual low-pass one; we give each output axis its own learned mix of the two, which is the most dangerous simple competitor. GSN-v, orbit counts concatenated to node features with ordinary propagation, tests whether structure helps as input rather than as a gate. IPGDN tests unsupervised channel disentanglement.
Three ablations isolate the key ingredient. The gate is driven by degree alone, by a learned per-node free parameter, and by orbit counts shuffled across nodes. All methods get the same node features, splits, depth grid (1 to 16 layers), and hyperparameter search budget of the same number of trials. The proposal beats the field only if it significantly raises second-axis correlation over the best competitor without lowering first-axis correlation. The significance test is a paired bootstrap over nodes across seeds. The orbit gate must also beat the degree-only gate. Beating a plain GCN alone counts for nothing.
Why not just…?
Why not just use a per-axis high-pass filterbank, which already keeps sharp differences between neighbors? This is the real threat. If the cross-cutting axis is simply rough everywhere, a per-channel low/high-pass mix recovers it, and orbits add nothing. Our hypothesis says roughness is concentrated at bridging nodes, so a node-level structural gate should beat a graph-level filter choice. The planted-structure synthetic experiment and the per-axis filterbank baseline settle this directly. If the filterbank matches the orbit gate on real graphs, we report the diagnosis as the main result and drop the orbit mechanism.
Why not just add PairNorm or skip connections? PairNorm keeps representations spread out in total. By its authors' own account in PairNorm, it prevents degradation with depth but does not raise peak accuracy. Nothing in it tells the model which direction of spread to keep. So it can maintain variance along the partisan axis while the cross-cutting axis still vanishes. The per-axis retention curves test this directly: if PairNorm flattens the second-axis decay as well as the gate does, the mechanism is unnecessary.
Why not just skip the graph and use a language model? On the first axis, text probes are already strong (Spearman about 0.87–0.88 in Linear Representations of Political Perspective). Text-only facet classifiers generalize poorly across topics, however. Ideology Takes Multiple Looks reports Macro-F1 around 42% on its 12-facet MITweet benchmark. Many online nodes also have sparse political text. The honest answer is empirical: if text-only probes match graph methods on the second axis, the graph matters only for the first axis. That outcome would itself be informative, because it would say graphs carry partisanship and little else.
Why not just use node degree instead of orbit counts? Sorted neighbor degrees are strong structural fingerprints (Deanonymizing Social Networks Using Structural Information). But degree does not separate a hub inside one bloc from a node bridging two blocs, and Motif-Role-Fingerprints shows role information is invisible to counts of connections. The degree-gate ablation decides this.
Proposed Experiments
The first testbed is synthetic. We generate graphs whose communities follow axis one while axis two is drawn independently, with a controllable fraction of bridging nodes carrying extreme axis-two values. Here we can verify the $(1-R_k)^L$ prediction and check whether the orbit gate recovers axis two only when axis-two signal is actually concentrated at bridges. A control condition spreads axis two uniformly, and there the gate should give no gain. That control keeps the method honest.
The main real testbed is 116th-Congress legislators with two-dimensional DW-NOMINATE targets, on the relational network of Legislator Representation Learning. We restrict it to non-vote relations so the targets do not leak into the graph. We also use the German Bundestag co-voting signed graph from Discovering Opinion Intervals. The second axis is derived from voting records held out from edge construction, and the interval method's 38% reduction in disagreements over correlation clustering serves as a structural reference. The third testbed is the 2020 election retweet-community network with ANES feeling-thermometer ground truth from Reading Between the Tweets. We project the 30 target stances onto their leading components. We note that the paper's community-specific target ranking reaches only mid-teens Spearman against high-40s for target-specific ranking. That gap fits our hypothesis that within-community, cross-cutting variation is what current propagation loses, and we will test whether the second component carries the gap. The Reddit politosphere of Modeling Ideological Agenda Setting and Framing (605 subreddits) serves only for the measurement study, since it lacks a multi-axis ground truth.
The metrics are per-axis Spearman correlation against ground truth as a function of depth, per-axis Dirichlet energy under the normalized Laplacian together with the Frobenius norm, and the ratio of second- to first-axis retention.
Three checks decide the project. First, in plain GCNs the second axis must decay measurably faster with depth than the first, and $R_k$ computed from the ground truth must predict the ordering. If not, the core claim is false. Second, the orbit gate must beat the per-axis filterbank, PairNorm and the degree gate on second-axis correlation at matched budget, while leaving first-axis correlation unchanged. Third, the learned second-axis gates must load on bridging orbits (path centers, clique peripheries) consistently across at least two of the three real networks. Without that consistency, the structural story is an overfit rather than a finding.
References
- PairNorm: Tackling Oversmoothing in GNNs
- Complete the Missing Half: Augmenting Aggregation Filtering with Diversification for Graph Convolutional Networks
- Surrogate explanations for role discovery on graphs
- Analysis of Dirichlet Energies as Over-smoothing Measures
- Are we measuring oversmoothing in graph neural networks correctly?
- Improving Graph Neural Network Expressivity via Subgraph Isomorphism Counting
- Independence Promoted Graph Disentangled Networks
- When Do GNNs Work: Understanding and Improving Neighborhood Aggregation
- Discovering Opinion Intervals from Conflicts in Signed Graphs
- Legislator Representation Learning with Social Context and Expert Knowledge
- Modeling Ideological Salience and Framing in Polarized Online Groups with Graph Neural Networks and Structured Sparsity
- Linear Representations of Political Perspective Emerge in Large Language Models
- Ideology Takes Multiple Looks: A High-Quality Dataset for Multifaceted Ideology Detection
- Deanonymizing Social Networks Using Structural Information
- Motif-Role-Fingerprints: The Building-Blocks of Motifs, Clustering-Coefficients and Transitivities in Directed Networks
- Reading Between the Tweets: Deciphering Ideological Stances of Interconnected Mixed-Ideology Communities