Many real graphs attach more than one kind of data to each node. A product in a co-purchase network has an image and a text description, and a paper in a citation network may have an abstract and a figure. Two things go wrong in such graphs, and they are usually studied separately. The class labels used for training can be wrong. One modality on a node can also be broken: a blurry image, an irrelevant caption, or a feature vector with inflated magnitude. Methods that make graph neural networks (models that update each node by aggregating its neighbors) robust to wrong labels treat each node's features as one trustworthy vector. Methods that guard multimodal models against a broken modality look at each sample alone and assume the labels are clean. When both problems occur together, each family misreads the other's failure. A node whose image is corrupted looks like a node whose label is wrong, and a wrong label corrupts the supervision that multimodal gates learn from.
This note proposes to separate the two failures by triangulation. Each node has three independent witnesses: its graph neighbors, its nearest neighbors in modality one, and its nearest neighbors in modality two. A wrong label disagrees with most witnesses at once. A broken modality produces a single witness that disagrees with the other two, whatever the label says. Counting who disagrees with whom gives two separate scores per node, one for label suspicion and one for corruption of each modality. These scores can then drive two different corrections: soft relabeling for suspect labels, and per-modality gating of message passing for suspect modalities. None of these scores is computed from the supervised loss.
Core Claim
On graphs whose nodes carry two or more modalities, label noise and modality corruption have different agreement signatures across the views "topology, modality 1, modality 2." A detector that reads these signatures will flag wrong labels more accurately than detectors that pool all features into one attribute view. It will also isolate corrupted modalities without any supervised gate. Using the two scores for different interventions should beat both concatenate-then-denoise and denoise-then-fuse pipelines when both kinds of noise are present. The mechanism is built from rank-based neighborhoods, so it is insensitive to feature scale by construction. That is the property evidential fusion work shows naive fusion lacks.
Concretely, for each view $v \in \{G, 1, 2\}$ and node $i$, let $p_i^v$ be the distribution of current (soft) labels among $i$'s neighbors in that view, excluding $i$ itself. For the graph view, neighbors are weighted by an approximate personalized PageRank. For a modality view, neighbors are the $k$ nearest nodes by cosine rank, so multiplying a modality by 100 does not change them. With $\tilde y_i$ the observed label, the label support from view $v$ is $p_i^v(\tilde y_i)$. The corruption score for modality $m$ at node $i$ measures how far its view departs from the consensus of the other two, discounted by how well those two agree with each other:
$$c_i^m = \mathrm{JS}\!\left(p_i^m,\ \tfrac{1}{2}(p_i^G + p_i^{m'})\right) - \mathrm{JS}\!\left(p_i^G,\ p_i^{m'}\right),$$
Here JS is Jensen–Shannon divergence and $m'$ is the other modality. The label is judged clean when at least two views support it and the dissenting view has a high corruption score. Following ICGNN, a two-component Gaussian mixture fitted to the combined support turns this into a posterior clean probability. The scores are recomputed each epoch as soft labels improve, much as DREAM refreshes its homogeneity weights.
How This Relates to Prior Work
Two graph label-noise methods motivate and supply the machinery. ICGNN flags a node only when its label contradicts both a structural diffusion and a feature-similarity graph, then corrects it softly. DREAM reweights nodes using embedding-proximity anchors and topology anchors. Both methods have a single attribute view. If one modality inside that view is corrupted, the attribute graph is polluted and the dual check loses its independence. We reuse their soft correction and mixture model. What changes is that the attribute view is split per modality and each modality is scored rather than trusted. NoisyGL constrains the design in three ways. It shows that message passing spreads errors, that i.i.d. label-noise methods such as Co-teaching fail on graphs, and that some robust methods run thousands of times slower than a graph convolutional network. We therefore use precomputed neighbor sets and polynomial approximations instead of matrix inversion.
The closest cross-direction work is IN²R. It fixes wrong image–caption pairings by trusting intra-modal nearest-neighbor graphs. It addresses correspondence noise rather than class-label noise, builds its graphs within each modality separately, and never tests whether a modality itself is lying. Our premise also partly contradicts its title. We assume intra-modal neighbors can lie when a modality is corrupted, and we use the graph and the other modality to catch them. The authors of IN²R also report an echo-chamber risk for rare classes, and triangulation gives a direct handle on it: a rare-class node whose graph neighbors disagree but whose two modalities agree keeps its label. Multimodal graph recommenders such as MENTOR and the modality-guided de-noising GNN filter noisy modality signals using interaction structure, but they assume the targets are clean.
From the multimodal side, Beyond Magnitude shows that sum, mean and Dempster–Shafer fusion drop to 0% accuracy when one view emits large, wrong logits. This motivates making every score rank-based. AFter and RSFNet show that per-sample routing and asymmetric gating prevent a bad modality from polluting a good one. However, their gates are trained on the task loss, which would teach them to fit wrong labels. Our gates are the label-free corruption scores.
Research Program
The model keeps one message-passing channel per modality over the shared graph. The message from neighbor $j$ in modality $m$ is scaled by $1 - \sigma(c_j^m)$, where $\sigma$ is a sigmoid, so a node's broken image stops propagating while its clean text still travels. At each node, the channel outputs are fused with weights from the node's own corruption scores rather than from a learned router. The loss is cross-entropy against a convex mix of the observed label and the neighbor-consensus label, weighted by the clean posterior. The scores are refreshed every epoch from the current soft labels.
The analysis part asks when triangulation is identifiable. It needs the three views' neighborhoods to overlap only partly. If the graph were simply the kNN graph of one modality, the views would not be independent witnesses. We will measure view-neighborhood overlap on each dataset and report detection quality as a function of it. We also expect failure where ICGNN reports it, around 80% noise, where all consensus is corrupted. We will locate that breakpoint rather than claim to move it.
Baselines
The naive reference is a vanilla graph convolutional network on concatenated modalities, trained directly on noisy labels. DREAM reports this family collapsing to 32.78% accuracy on Cora with 70% uniform noise, against 60.22% for DREAM. That gap shows how much a graph-aware defense is worth in the unimodal case.
The strongest existing competitors are ICGNN and DREAM run on concatenated, per-modality z-scored features. ICGNN reports sustaining over 80% accuracy on DBLP with 40% pair noise. These are the right opponents because their core idea, checking a label against structure and attributes, is exactly what we refine. Z-scoring each modality before concatenation gives them the cheapest available fix for scale hijacking.
A second strong baseline runs late fusion. It trains a separate ICGNN per modality and averages the predictions, and we also test a variant that averages with SAEF's scale-invariant fusion from Beyond Magnitude. A third is the two-stage pipeline: an AFter-style router trained on noisy labels, followed by a robust GNN.
All methods share data splits, noise seeds, backbone width, and a hyperparameter search of equal trial budget chosen on a small held-out set with noisy labels. We report wall-clock time relative to a graph convolutional network, because NoisyGL documents large runtime gaps.
Two ablations isolate the key ingredient. The first uses the same model but computes label suspicion from one pooled attribute view, so triangulation is removed. The second keeps the triangulated label scores but removes per-modality gating of messages. Beating the baselines means higher test accuracy averaged over noise types in the joint-noise regime, together with no loss beyond noise-seed variance when only label noise is present. It also means a higher AUROC for detecting wrong labels. No reported result was found on Lacuna for any method on node classification with simultaneous label noise and modality corruption, so all numbers in this regime will be new.
Why not just…?
Why not just z-score each modality, concatenate, and run DREAM or ICGNN? Normalization fixes scale hijacking but not content corruption. A node whose image features are replaced with another node's image still has normally scaled features that point to the wrong neighbors. In ICGNN, that pollutes the attribute contradiction check. This is a real threat when corruption is pure scale inflation, and we expect it to be competitive there. The swap-corruption condition below is designed to separate the two approaches, and we will report both conditions.
Why not just train a robust GNN per modality and average them? This is the most serious alternative, because it already stops cross-modal contamination inside the network. Its weakness is that each per-modality model decides label cleanliness from its own view, so it cannot tell "my view is broken here" from "this label is wrong." Averaging also dilutes a correct view with a broken one on exactly the nodes where corruption is concentrated. If late fusion with SAEF matches our method on label-noise detection AUROC and accuracy, then triangulation adds nothing, and we will say so.
Why not just drop the noisier modality? When corruption is spread across nodes rather than confined to one modality, each modality is clean on most nodes. Dropping either one throws away the clean signal on the nodes where the other is broken. The unimodal robust GNN on each modality will be reported as a reference, and our method must beat the better of the two.
Proposed Experiments
The file names no labeled multimodal node-classification benchmark with label noise, so we use two testbeds. The first is semi-synthetic. On Cora, CiteSeer, PubMed, Amazon-Photo, DBLP and OGBN-arxiv, we split node features into two disjoint "modalities," either by feature block or by pairing text features with a second embedding from a different encoder. Label noise follows NoisyGL's uniform and pair protocols plus an instance-dependent variant, at rates from 20% to 80%.
Modality corruption is injected on 10–40% of nodes, independently of the labels, in three forms. The first adds Gaussian noise. The second inflates magnitude by up to 100×, mirroring the scale stress test in Beyond Magnitude. The third swaps a node's modality with a random other node's, a feature-level analogue of that paper's lying-logit test. Cornell and Flickr test the heterophilous regime, where the graph witness is weak. There, triangulation should fall back to the two modalities, and we will check whether it does.
The second testbed is real multimodal: item graphs with image and text features from the multimodal recommendation setting of MENTOR, with item category as the label and synthetic label noise. Metrics are test accuracy, AUROC for detecting wrong labels, AUROC for detecting corrupted modalities against the injected ground truth, per-class recall on the rarest classes, and runtime.
Three checks decide the outcome. First, in the joint-noise swap condition, the triangulated detector must beat both the pooled-view ablation and per-modality late fusion on label-noise AUROC. If it does not, the core claim fails. Second, modality-corruption AUROC must be well above chance without any label supervision, and it must stay high as label noise rises to 60%. This shows that the label-free gate really is decoupled from bad labels. Third, rare-class recall should degrade less than under ICGNN, which would show that disagreement among the three views protects minority nodes from the echo chamber identified in IN²R.
References
- Identifying and Correcting Label Noise for Robust GNNs via Influence Contradiction
- DREAM: Dual-Standard Semantic Homogeneity with Dynamic Optimization for Graph Learning with Label Noise
- NoisyGL: A Comprehensive Benchmark for Graph Neural Networks under Label Noise
- Intra-Modal Neighbors Never Lie: Rectifying Inter-Modal Noisy Correspondence via Graph-Based Intra-Modal Reasoning
- MENTOR: Multi-level Self-supervised Learning for Multimodal Recommendation
- Multimodal Graph Neural Network for Recommendation with Dynamic De-redundancy and Modality-Guided Feature De-noisy
- Beyond Magnitude: Scale-Invariant Evidential Fusion for Multi-View Classification
- AFTER: Attention-Based Fusion Router for RGBT Tracking
- Residual spatial fusion network for RGB-thermal semantic segmentation