Many prediction models for tabular and recommendation data now treat each input field (a user's city, the app category, the hour of day) as a node in a graph. They learn which pairs of fields should interact, which amounts to deciding which edges to keep. The best of these models choose edges separately for each example, scoring each candidate pair from the two fields' current values. That design has a weakness nobody has measured. When a field's value is wrong because it was mistyped, defaulted, dropped in transmission, or hashed into the wrong bucket, the wrong value does two kinds of damage. It enters the prediction directly, and it also rewires the graph, so one bad field can change which other fields talk to each other. Sparse mixture-of-experts models show the same failure, where noise that fools the routing function sends inputs to the wrong sub-network (Robustness of Mixtures of Experts to Feature Noise).
We propose to split the edge decision into two separate modules with deliberately restricted inputs. The first is a topology module that decides which fields may interact, using only field identity and coarse context, never the possibly corrupted values. The second is a reliability module that estimates, for each field in each example, how much its value can be trusted. It judges this mainly from whether the value agrees with the other fields. The reliability score then gates the graph in one direction only. A field judged unreliable stops sending messages to its neighbors, but it keeps receiving them, so the graph stays connected and a corrupted field can still be partly reconstructed from its context. The question is whether this split degrades more slowly under field-level corruption than existing per-example edge selectors trained with the same noise augmentation, and whether it does so at no cost on clean data.
Core Claim
Existing work on noise in feature graphs means something narrower than corrupted inputs. Detecting Beneficial Feature Interactions for Recommender Systems (L0-SIGN) prunes edges with a differentiable on/off gate and shows that pruning raises Frappe AUC from 0.9448 to 0.9580. Its "noise," however, is useless pairs within clean inputs. GraphFM selects edges per layer with a learned pair scorer and shows that removing selection hurts on Criteo, Avazu and MovieLens-1M, but it never corrupts inputs. Fi-GNN keeps all edges with attention weights computed from the embeddings. In all three models, edges are scored from the very embeddings that corruption damages.
Our claim is that, under field corruption, edge selection driven by values is a liability. We predict that a value-blind topology combined with a separate, sender-side reliability gate keeps most of the clean-data benefit of selective interaction while removing the "rewiring" channel of damage. The new element is not adding a denoiser to a feature graph. It is the specific separation: topology is decided without values, trust is decided per example, and trust is applied asymmetrically to outgoing messages. That asymmetry is our answer to the tension between pruning and connectivity raised in the research file. Fully pruning a noisy node isolates it. Muting only its outgoing edges blocks the spread of noise without fragmenting the graph.
Concretely, one message-passing step for field $i$ becomes
$$h_i' = h_i + \sum_{j \neq i} a_{ij}\, r_j\, W_{ij} h_j ,$$
where $h_i$ is field $i$'s embedding, $a_{ij}\in[0,1]$ is the edge weight from the topology module (a function of field identities only), $r_j\in[0,1]$ is the reliability of field $j$ in this example, and $W_{ij}$ is an edge-specific transform as in Fi-GNN. The reliability module is trained with an auxiliary detection loss. During training a corruption module perturbs a random subset of fields, and the reliability module must predict which fields were perturbed:
$$\mathcal{L} = \mathcal{L}_{\text{task}} + \lambda \sum_j \text{BCE}\big(r_j,\ 1-c_j\big),$$
where $c_j\in\{0,1\}$ marks whether field $j$ was corrupted and $\lambda$ weights the auxiliary term. This borrows the idea of self-supervised distortion prediction used for trajectories in Enhancing Trajectory Prediction through Self-Supervised Waypoint Distortion Prediction, applied to field-level trust instead of coordinate recovery.
Research Program
The work has three parts. The first part is a harness for controlled comparison, built on the decoupling pattern of UCEE, which separates dataset generation from estimators behind a fixed interface, and of FaceX-Zoo, which separates backbones from training heads. The harness has four swappable parts with fixed contracts: a corruption module, an edge selector, an aggregator and a training objective. The edge selectors are the L0 hard-concrete gate, a GraphFM-style pair scorer, top-k selection, a static learned graph, and our value-blind topology. The aggregators are a gated recurrent update as in Fi-GNN and multi-head attention. The harness is a means, not the contribution. It exists so that every result below differs from its comparison by exactly one part, which addresses the concern that gains come from implementation drift (DeepDIVA). We accept the cost the file notes, that abstract boundaries block kernel fusion. Wall-clock time is therefore reported only within the harness.
The second part is a measurement study that comes before any new model. We measure topology drift: for each per-example selector, how much the chosen edge set changes (Jaccard distance between the edges chosen on clean and corrupted copies of the same example) as a function of corruption rate. We also measure how much of the AUC loss is explained by that drift. To do this we run a "frozen-topology" variant that uses the clean example's edges with corrupted values, and compare it with the variant where both edges and values come from the corrupted input. If drift explains little of the loss, the main premise is wrong, and we will report that outcome instead of the method.
The third part is the method above, together with its failure analysis. The reliability module detects corruption through cross-field consistency, so it should fail once too many fields are corrupted at the same time. This is the same breakdown as router collapse in Robustness of Mixtures of Experts to Feature Noise. We will locate that breakdown point empirically and not claim robustness beyond it.
Baselines
The naive reference point is each graph model trained and tested exactly as published, on clean data, then evaluated on corrupted test sets. It shows how large the problem is.
The strongest competitor is a per-example edge selector trained with the same corruption augmentation that our model sees. We use L0-SIGN on Frappe, where it reports AUC 0.9580 on clean data (Detecting Beneficial Feature Interactions for Recommender Systems), and GraphFM and Fi-GNN on Criteo, where Fi-GNN reports AUC 0.8062 (Fi-GNN). For GraphFM, Lacuna's summary gives no numeric result on Criteo, Avazu or MovieLens-1M, so we use our own reproduction. No reported result was found on Lacuna for any of these models under input corruption, so every corrupted-data number will come from our runs. Augmented L0-SIGN is the right opponent because it already learns sparse, hard-zero edges and, once trained on noise, could learn to route around bad fields on its own. If augmentation alone fixes the problem, our separation is unnecessary. Fairness rules apply to every model in the comparison. They share the same harness, embedding size, corruption schedule and epochs, use the same number of hyperparameter trials (for example, 30 per model drawn from equal-width search spaces), and run five seeds each. A first sanity check is that the reproductions must land near the published clean numbers.
The third baseline is a static learned graph. It is value-blind like ours but has no reliability gate, so it tests whether removing values from edge selection is enough by itself.
The ablations isolate the key ingredient. One adds the reliability gate to the per-example selectors, which separates the effect of trust from the effect of value-blind topology. One applies the gate symmetrically, muting both incoming and outgoing edges, which tests the asymmetry claim. One drops the detection loss, so that $r_j$ is learned only from the task.
Beating the strongest baseline means two things. At 30% field corruption, the AUC gap must exceed twice the seed standard deviation on at least three of the four datasets. And clean AUC must not fall by more than 0.001, since gaps of that size matter in click prediction.
Why not just…?
Why not just train the existing models with noise augmentation? This is the most serious threat. It is plausible that augmented L0-SIGN learns to ignore unreliable fields, and so it is our main baseline rather than a straw man. The frozen-topology diagnostic settles it. If augmented models show little topology drift, or if drift does not cost accuracy, then the value-driven rewiring channel is already handled and our claim fails.
Why not just impute or flag missing values upstream and keep the model unchanged? Upstream cleaning works when corruption is visible, as with an empty field. It cannot help when a field holds a plausible but wrong category. MVB-STNet found that deterministic pipelines degraded under 30% sensor noise, while uncertainty built into the model cut RMSE by 10.9% on PeMS08, but that evidence comes from traffic sensors, not field graphs. We will therefore include an imputation-plus-missingness-indicator baseline. We expect it to compete on the dropout corruption and lose on the substitution corruption. If it wins on substitution too, we will say so.
Why not just use a fixed global graph, or a dense model with Jacobian regularization? The static graph is already a baseline. A Lipschitz penalty of the kind used in Denoising Self-Attentive Sequential Recommendation bounds sensitivity to small perturbations. Categorical substitution, however, is not a small perturbation in embedding space: it swaps one embedding for an unrelated one. We include Jacobian-regularized Fi-GNN to test this.
Proposed Experiments
The datasets are Criteo, Avazu, Frappe and MovieLens-1M, split 8:1:1 as in GraphFM. Corruption comes in three types: substitution of a categorical field's value with a draw from that field's overall value distribution, dropping a field to its default token, and Gaussian noise on Criteo's numeric fields after log transform. Each type is applied at 0% to 50% of fields per example. As an extra check, we also evaluate on Criteo examples whose fields are naturally empty. We report AUC and LogLoss, area under the AUC-versus-corruption curve, topology drift, and how well the reliability module detects corrupted fields, measured as AUC against the corruption labels.
Four checks decide the outcome. The first is the frozen-topology diagnostic described above. The second is the curve comparison against augmented L0-SIGN and GraphFM. The third is the symmetric-versus-sender-only gating ablation at high corruption, where the asymmetry argument predicts the largest difference. The fourth is a transfer test: train with substitution and test with dropout, and the reverse. This shows whether the reliability module has learned cross-field consistency or merely memorized one corruption type. If the reliability module's gains disappear under transfer, the honest conclusion is that value-blind topology helps and learned trust does not, and the note will be revised around that narrower result.
References
- Robustness of Mixtures of Experts to Feature Noise
- Detecting Beneficial Feature Interactions for Recommender Systems
- GraphFM: Graph Factorization Machines for Feature Interaction Modeling
- Fi-GNN: Modeling Feature Interactions via Graph Neural Networks for CTR Prediction
- Enhancing Trajectory Prediction through Self-Supervised Waypoint Distortion Prediction
- Universal Causal Evaluation Engine: An API for empirically evaluating causal inference models
- FaceX-Zoo: A PyTorch Toolbox for Face Recognition
- DeepDIVA: A Highly-Functional Python Framework for Reproducible Experiments
- Multi-view Bayesian spatio-temporal graph neural networks for reliable traffic flow prediction
- Denoising Self-attentive Sequential Recommendation