Graph fraud detectors are trained and scored on labels that some earlier system produced. On review platforms, the "fraudulent" tag usually means "flagged by the platform's own filter." A model that scores well on such data may be learning to imitate that filter, not to find fraud. Clinical researchers know this trap as retrospective selection bias: a historical cohort contains only the patients that past decisions let in, so models learn those decisions. This note proposes treating fraud labels in the same way. A label is the outcome of a selection process. Its probability depends on where an account sits in the graph and on how loudly it behaves. We correct for that process using a small audit of truly random nodes, and we evaluate detectors on the fraudsters the historical filter missed.
The question has an immediate scientific payoff. The literature carries an unresolved contradiction. Fraud studies report that fraudsters sit at the structural margins, with low PageRank and little overlap between their neighbors across relation types (Don't Ignore Alienation and Marginalization: Correlating Fraud Detection). Hate-speech studies report that malicious users move quickly into the dense core of the network (Hate begets Hate: A Temporal Study of Hate Speech). Both findings are measured only on actors that some filter caught or retained. If the filter's catch rate depends on network position, the "structural signature of maliciousness" may partly be a signature of the filter. We propose to test this directly, in a setting where the true answer is known.
Core Claim
In graph fraud data, the chance that a real fraudster ends up labeled depends on its structural position and its behavioral loudness. This dependence makes current detectors, and the published structural signatures of fraud, biased toward fraud that is easy to catch. A position-aware correction calibrated on a small random audit recovers the missed, marginal fraudsters better than three alternatives: debiasing that ignores the selection mechanism, degree-bias repair, or simply training on the audit.
The formal object is a positive-unlabeled setup with a selection propensity that depends on graph position. Let $y$ be true fraud status, $s$ the observed label (1 if the historical filter flagged the node), $x$ node attributes and $g$ structural position (degree, PageRank, $k$-core depth, cross-relation overlap). Assume the filter flags only real fraud. Then
$$P(s=1 \mid x, g) = e(x, g)\, P(y=1 \mid x, g), \qquad e(x,g) = P(s=1 \mid y=1, x, g),$$
so a detector trained on $s$ estimates fraud risk multiplied by the filter's catch rate $e$. If $e$ is high for loud, central fraudsters and low for quiet, marginal ones, the bias falls exactly on the marginal ones.
The ratio $P(s{=}1 \mid x,g)/e(x,g)$ recovers true risk. Estimating $e$ from labels alone is not identified, because the observed labels cannot separate "rarely fraud" from "rarely caught." We restore identification with a small, uniformly random audit set whose true labels are known. Industrial trust-and-safety teams already maintain such manual-review samples.
What is new relative to existing work is the explicit model of $e$ as a function of graph position. Debiased Graph Neural Networks With Agnostic Label Selection Bias also treats labeled nodes as a non-random sample, but it handles this by decorrelating embedding dimensions without modeling the mechanism. It assumes the topology is fixed and innocent, and it targets citation-graph classification rather than one-sided fraud labels.
Research Program
The first component is a controlled testbed where true labels are complete. We take graphs with full ground truth or near-complete labels and hide positives according to simulated historical filters. The anchor is G-Internet, a synthesized benchmark generated from 12 known fraud rules, used by Association-Focused Path Aggregation for Graph Fraud Detection with code at horrible-dong/gpa. We add T-Finance and T-Social, and we treat the existing YelpChi and Amazon labels as "truth" for re-filtering.
We define three filter families. The first is a burst filter modeled on the behavioral cues that What Yelp Fake Review Filter Might Be Doing? found Yelp's filter relies on, such as too many reviews in a day and extreme ratings. The second is a centrality filter that preferentially catches high-PageRank or high-$k$-core fraudsters. The third is a uniform filter as a control. Each filter has a strength parameter that controls how unequal catch rates are across positions.
The second component is the estimator. We fit a propensity model $e(x,g)$ as a small graph-feature classifier on the audit set's true positives, predicting which of them the filter flagged. We then train the detector with inverse-propensity-weighted positive-unlabeled loss, clipping weights to control variance on very marginal nodes. The detector backbone is held fixed across all methods. We use Partitioning Message Passing (Partitioning Message Passing for Graph Fraud Detection), which already handles fraudsters hiding among benign neighbors. This makes the contribution the correction itself, not a new architecture.
The third component is a signature audit. Using the same pipeline, we measure the PageRank and cross-relation-overlap gap between fraud and benign nodes three times: on filtered labels, on corrected estimates, and on true labels. This tests whether marginalization of the kind reported for COFRAUD is a property of fraud or of the filter. A short companion analysis applies the same survivorship logic to the Gab core-migration finding by simulating early suspension of peripheral accounts on the Gab temporal corpus. That analysis is secondary and only needs to show whether a 3.36- versus 6.94-month gap could arise from attrition alone.
Baselines
The naive reference point is a detector trained directly on the filtered labels as though they were truth, which is the current benchmark practice. We run the same PMP backbone, the same features and the same hyperparameter budget. Any method that does not beat it on hidden-fraud recall is pointless.
The strongest existing detectors are SCSE-GFD (Spectrum-Constrained and Skip-Enhanced Graph Fraud Detection: Addressing Heterophily in Fraud Detection with Spectral and Spatial Modeling) and GPA. SCSE-GFD reports an AUC of 90.58% and F1-macro of 76.12% on YelpChi, and an AUC of 96.21% on Amazon. GPA reports a +15.7% relative improvement in Average Precision over the best baseline on G-Internet. The file gives no absolute value. These are the right competitors because their heterophily and multi-hop path machinery could plausibly find quiet fraudsters without any explicit correction. We run each with its released settings and an equal tuning budget, selected on audit-set validation. Beating them means higher recall at fixed precision on the hidden positives, the fraudsters the filter missed. Overall AUC against filtered labels does not count, because it rewards imitating the filter. Under our relabeling protocol, there is no reported result found on Lacuna for any of these methods.
The mechanism-agnostic debiasing baseline is DGNN's differentiated variable decorrelation applied on top of the same backbone. DGNN reports gains of up to 17.0% over GCN/GAT under heavy label selection bias on citation graphs. The degree-repair baseline is L2DGCN (L2DGCN: Learnable Enhancement and Label Selection Dynamic Graph Convolutional Networks for Mitigating Degree Bias), which prunes hub edges and adds feature-similarity edges to low-degree nodes. A standard positive-unlabeled baseline uses a single constant propensity, which tests whether position dependence matters at all.
The key ablation removes $g$ from the propensity model, so that $e$ depends on attributes only. This isolates whether structural position carries selection information that attributes do not.
Why not just…?
Why not just train on the audit set, since it is unbiased? This is the most serious threat. With fraud rates in the low single digits and audits of 1–2% of nodes, the audit contains very few positives, but at large audit sizes it may win outright. We will sweep audit size and report the crossover point. Combining the large biased label set with a small unbiased one is only worthwhile below that crossover, and we will say so if the crossover is low.
Why not just use a strong heterophily-aware detector, which already looks past camouflage? Methods like PMP, SCSE-GFD and TAM (Truncated Affinity Maximization: One-class Homophily Modeling for Graph Anomaly Detection) fix how a node aggregates its neighbors. They do not fix which nodes the training signal calls positive. A detector that is excellent at learning the label distribution will learn the filter more faithfully. The experiments test this directly: under centrality filters, the strongest detectors should show the largest gap between AUC on filtered labels and recall on hidden positives. If they do not, the correction is unnecessary, and we will report that.
Why not just apply agnostic reweighting like DGNN? Decorrelating embeddings removes spurious correlations among features. One-sided selection is different: unlabeled fraudsters are systematically mislabeled as benign. Reweighting labeled nodes cannot turn any of them into positives. The constant-propensity and attribute-only ablations show how much of DGNN's shortfall comes from ignoring position.
Why not just repair degree bias, since marginal nodes are low-degree? L2DGCN treats low degree as missing information and fills it with nearest-neighbor edges based on feature similarity. When fraudsters copy benign attributes, as TAM documents, those added edges tie them to benign neighbors and dilute their signal further. Its results on marginal hidden fraud will show whether this concern holds.
Proposed Experiments and Decisive Checks
The main experiment crosses five graphs, three filter families and four filter strengths with audit sizes from 0.5% to 5%. All methods report precision-recall AUC and recall at fixed precision on hidden positives, broken down by PageRank quintile, with five seeds.
The first decisive check is the centrality filter at high strength. There, the correction should raise hidden-fraud recall in the bottom PageRank quintile without losing ground in the top quintile. Under the uniform filter it should match the naive baseline, because a correction that helps when there is no bias signals overfitting.
The second decisive check is the signature audit. If the fraud–benign PageRank gap measured on filtered labels moves substantially toward the true-label gap after correction, structural signatures reported on filtered benchmarks need re-examination. If the gap is unchanged under every filter, marginalization is robust and our correction is a recall tool rather than a scientific corrective.
The third check is a real-data plausibility test on YelpChi, where true labels are unavailable. We ask whether nodes that the corrected detector scores highly but that the filter left unlabeled are concentrated at the margins. We then hand-inspect a stratified sample against the behavioral cues from the Yelp filter study. This check is suggestive only and will be reported as such.
References
- Don't Ignore Alienation and Marginalization: Correlating Fraud Detection
- Hate begets Hate: A Temporal Study of Hate Speech
- Debiased Graph Neural Networks With Agnostic Label Selection Bias
- Association-Focused Path Aggregation for Graph Fraud Detection
- What Yelp Fake Review Filter Might Be Doing?
- Partitioning Message Passing for Graph Fraud Detection
- Spectrum-Constrained and Skip-Enhanced Graph Fraud Detection: Addressing Heterophily in Fraud Detection with Spectral and Spatial Modeling
- L2DGCN: Learnable Enhancement and Label Selection Dynamic Graph Convolutional Networks for Mitigating Degree Bias
- Truncated Affinity Maximization: One-class Homophily Modeling for Graph Anomaly Detection