Graph neural networks classify a node by averaging information from its neighbors. That works when neighbors tend to share the node's class, a property called homophily. It fails in fraud and transaction networks, where bad actors deliberately connect to good ones and the illicit signal gets averaged away. One family of fixes groups each node's neighbors by their (predicted) class and summarizes each group separately. The summary then records who a node talks to, not just an average of everyone it talks to. These label-wise methods are among the strongest on heterophilic benchmarks. Every one of them, however, assumes a frozen graph whose class predictions are computed once. Real fraud graphs stream in over time, the predicted labels of past neighbors are often wrong, and those predictions get corrected later.
This note proposes treating label-wise neighbor summaries as revisable running sums. When an edge arrives or a neighbor's predicted class changes, only the affected summaries are updated, by a small exact correction. Tracking how the class-conditional summaries shift between time windows also gives a cheap alarm for abrupt regime changes, such as the darknet-market shutdown that breaks every model on the Elliptic Bitcoin graph. The central question is empirical and falsifiable. Do revisable label-wise statistics, maintained online, beat both temporal GNNs and strong per-snapshot recomputation on evolving heterophilic graphs? And is the gain due to the revision and drift mechanisms rather than to the predictor placed on top?
Core Claim
Static label-wise methods motivate the proposal by showing that class-partitioned aggregation is what makes heterophily usable. LW-GCN aggregates neighbors separately per pseudo-label and reaches 86.2% accuracy on Texas, against 57.6% for a standard graph convolutional network. GAGA pools fraud-graph neighbors into fraudulent, benign and unknown groups and precomputes these group vectors so that it scales to 13 million nodes. HiGNN and HDP build per-node neighbor-label distribution profiles; HDP reaches 62.07% on Squirrel. LACA gates neighbor features by predicted fraud probability.
All five compute pseudo-labels and group aggregates once over a static adjacency matrix. Dynamic methods have the reverse gap. EvolveGCN evolves the network weights through time with a recurrent unit but averages neighbors uniformly within each snapshot. THeGCN learns signed temporal attention from features and time encodings and never uses class assignments. That distinction matters when fraudsters copy benign features, which is exactly the regime where LACA and LW-GCN argue feature similarity fails.
The claim is that the right object to carry through time is a set of soft label-wise sufficient statistics. For node $v$ and class $c$, let $p_u(c)$ be the current predicted probability that neighbor $u$ has class $c$ (or the true label, once revealed), and $x_u$ its features. The statistics are
$$m_{v,c} = \sum_{u \in N(v)} p_u(c)\, x_u, \qquad n_{v,c} = \sum_{u \in N(v)} p_u(c).$$
Because both are linear in $p_u$, any change can be applied exactly to every neighbor of $u$ as a small correction:
$$m_{v,c} \leftarrow m_{v,c} + \Delta p_u(c)\, x_u .$$
This covers a newly arrived edge, a revealed label, or a revised prediction for $u$. It avoids recomputing over the full graph and it makes pseudo-label errors correctable after the fact, which no static label-wise method allows.
The two-sided analysis of GGCN shows that over-smoothing in deep networks behaves like "pseudo-heterophily": misclassified nodes act as wrong-class neighbors and spread their errors across layers. We expect the streaming analog to be a temporal cascade. Wrong pseudo-labels at time $t$ corrupt the summaries used at $t+1$, which produce more wrong pseudo-labels. Revision is the proposed brake on that cascade.
Research Program
The system has three parts. First, a store of $m_{v,c}$ and $n_{v,c}$ for one and two hops, kept in separate class buckets rather than stacked, updated incrementally on edge and label events. Second, a light predictor that consumes a node's own features together with the normalized bucket means and counts. We will test both a small network and a gradient-boosted tree, because the statistic, not the predictor, is the hypothesized contribution. Third, a drift monitor. It compares, window to window, the distribution of class-conditional neighbor profiles, meaning the vector of $n_{v,c}$ normalized over $c$ and averaged within each predicted class. When the divergence exceeds a threshold calibrated on early windows, the monitor resets low-confidence pseudo-labels to uniform and briefly reweights recent labeled windows.
HiGNN constrains this design. It proves that profiles lose discriminative power when neighbor classes become near-uniform, its "mid-homophily pitfall." We therefore expect the drift alarm to fire exactly when profiles drift toward uniform, and we will report whether this coincides with the Elliptic time-step-43 collapse documented in the Elliptic paper.
Over-smoothing enters as a constraint and a measurement question, not as the headline. HDP propagates distribution profiles along trusted edges, which raises an open question: do label-wise profiles themselves homogenize across hops or across time steps? We will measure the variance of class-conditional profiles as hop depth and stream length grow. Following the critique of oversmoothing metrics, we will separate convergence within the same class (benign) from collapse across classes (pathological), rather than reporting a single average distance.
Baselines
The naive reference is a feature-only model applied per time step, trained on the same temporal split. On Elliptic, the Elliptic paper reports logistic regression at 0.481 illicit F1. It also reports Random Forest with all features, including its 72 hand-built neighbor aggregates, at 0.788 illicit F1. That Random Forest number is the strongest reported result in the file and the bar the method must clear. The same paper reports EvolveGCN at 0.720 and Skip-GCN at 0.705. We will reproduce Random Forest and a gradient-boosted tree with identical features and a matched tuning budget. Clearing 0.788 on the standard temporal test steps, with the gain concentrated after step 43, would count as beating it.
The strongest graph competitors are static label-wise methods recomputed from scratch at every snapshot: GAGA-style group aggregation and LW-GCN, retrained on all labels revealed so far. This is the fairest test of whether incremental revision matters, as opposed to label-wise aggregation alone. They receive the same features, the same label-reveal schedule and the same hyperparameter search. We will report both accuracy and the wall-clock cost of each update.
The temporal competitors are EvolveGCN-O and THeGCN. They are run on Elliptic and on THeGCN's traffic benchmarks, where THeGCN reports 55.82% accuracy on PEMSLA against 48.27% for its best temporal baseline. Beating THeGCN on PEMSLA, or matching it while using class buckets instead of feature-derived signs, would support the claim that categorical routing transfers to dynamic heterophily. Exceeding 55.82% is the concrete bar.
Two ablations isolate the key ingredients. The first freezes pseudo-labels once assigned (no revision); this tests the error-cascade hypothesis. The second disables the drift reset. A third control feeds our exact statistics to a gradient-boosted tree instead of the network. If the tree matches the network, we will say so: the contribution is then the statistic and its maintenance, not a new architecture.
Why not just…?
Why not just give a Random Forest the GAGA-style group features, recomputed every snapshot? This is the most serious threat. Trees already beat every GNN on Elliptic (0.788 versus 0.720 for EvolveGCN, per the Elliptic paper), and GAGA's group vectors are precomputable. We have reshaped the proposal around this alternative rather than against it: the tree-on-our-statistics arm is part of the method. What recomputation lacks is revision. Each snapshot's groups are built from whatever pseudo-labels existed then, and nothing propagates later corrections. If per-snapshot recomputation with true labels-to-date matches our revisable version both before and after step 43, the revision claim is refuted. What would remain is a cost argument, and that argument is weak on a graph of Elliptic's size.
Why not just add a high-pass filter channel, as FB-GNN does? FB-GNN adds a channel that computes differences between a node and its neighbors, roughly 10 points above the standard network on Texas and Cornell. Label-wise methods report about 86% there against the standard network's 57.6%. FB-GNN also depends on a static Laplacian, the matrix encoding the graph's connectivity. We will include a temporal filterbank variant on Elliptic to check whether this gap persists on transaction data.
Why not just let THeGCN's signed attention learn which neighbors to subtract? It may, when features are honest. The camouflage argument from LACA and LW-GCN predicts it fails when illicit nodes mimic benign features. Two checks will settle this. We will compare the methods on the illicit class specifically. We will also run a controlled synthetic stream, built from a stochastic block model as in EvolveGCN's benchmarks, in which minority-class features are progressively interpolated toward the majority while their connection pattern is kept fixed. Our method should degrade more slowly than THeGCN as camouflage increases; if it does not, categorical routing adds nothing over learned signs.
Why not just retrain EvolveGCN after the shock? Retraining needs labels from after the shock, and those labels arrive late. The drift monitor needs no labels. The relevant measurement is detection lag versus label-reveal lag around step 43.
Proposed Experiments
The primary benchmark is Elliptic under its original temporal train/test split, reporting illicit F1 per time step as well as aggregated, with separate means before and after step 43. Secondary benchmarks are THeGCN's PEMSLA, PEMSBA, PEMSSD, Reddit and Brain streams, plus the synthetic stochastic block model streams with injected label-mixing shocks of known time and size.
The decisive checks are four. First, the frozen-pseudo-label ablation should show a widening gap over time, measurable as the growth of the error rate among nodes whose neighbors were mispredicted; this is the signature of the temporal cascade. Second, on synthetic streams, the drift alarm should fire within a window or two of the injected shock and, with resets, recover F1 faster than without. On Elliptic, it should flag step 43 before labels reveal the collapse. Third, cross-class profile variance should stay bounded as hops and stream length grow; if it collapses, the "distributional over-smoothing" concern from HDP is confirmed, and the separate-bucket design must be revisited. Fourth, per-event update cost should stay roughly constant as the stream grows, while full recomputation cost grows with graph size.
Any of these failing is informative. In particular, if the tree baseline on recomputed group features ties us everywhere, the honest conclusion is that label-wise grouping, not online revision, is what dynamic fraud detection needs.
References
- Label-Wise Graph Convolutional Network for Heterophilic Graphs
- Label Information Enhanced Fraud Detection against Low Homophily in Graphs
- Learn from Heterophily: Heterophilous Information-enhanced Graph Neural Network
- Heterophilous distribution propagation for Graph Neural Networks
- Clustering-Augmented Fraud Detection on Graphs Using Label-Aware Feature Aggregation
- EvolveGCN: Evolving Graph Convolutional Networks for Dynamic Graphs
- THeGCN: Temporal Heterophilic Graph Convolutional Network
- Two Sides of the Same Coin: Heterophily and Oversmoothing in Graph Convolutional Neural Networks
- Anti-Money Laundering in Bitcoin: Experimenting with Graph Convolutional Networks for Financial Forensics
- Are we measuring oversmoothing in graph neural networks correctly?
- Complete the Missing Half: Augmenting Aggregation Filtering with Diversification for Graph Convolutional Networks