Anomaly detectors for evolving graphs, such as fraud monitors on transaction networks, are trained on one period or platform and then deployed on data that drifts. The standard fix is test-time adaptation: nudging the model on each new batch of unlabeled data as it arrives. This proposal argues that the usual recipe for this undermines anomaly detection in particular. Current methods decide which predictions to trust by checking whether a node agrees with its neighbors. They then keep pulling representations toward the prototypes of the dominant "normal" class. Anomalies, however, are by definition the nodes that disagree with their neighbors and with their own history. Repeated adaptation therefore behaves like the familiar over-smoothing of deep graph networks, where node representations blur into each other, except that the blurring happens across adaptation steps rather than across layers.
We propose to adapt the detector with a self-supervised consistency signal built from semantic-preserving augmentations, meaning perturbed copies of the input designed not to change what makes a node normal or anomalous. Concretely, each view perturbs only the smooth, history-consistent parts of the data: edges between similar nodes, and the component of a node's features that its past already explains. It never touches the local and temporal deviations that carry the anomaly signal. The adaptation is guarded by a cheap rank-based collapse monitor that rejects any update which makes the representation space degenerate. The two-sentence version: adapt using only the parts of the graph that cannot contain the anomaly, and refuse updates that shrink the representation space.
Core Claim
The closest existing system, OTTA-DGAD, adapts a pretrained graph-network-plus-Transformer detector to sequential unlabeled chunks. It does this through pseudo-labels chosen by an "entropy-structural reliability score", which trusts a prediction only when it is confident and agrees with the local neighborhood. It also uses prototype memory and a cross-chunk context buffer. That neighbor-agreement rule sits in direct tension with the premise of local-inconsistency anomaly detection that GADAM builds on. GADAM computes anomaly scores with a plain feature network precisely because neighborhood aggregation washes out anomalies that differ from their surroundings.
We claim three things, each testable. First, pseudo-label adaptation of this kind systematically loses recall on the most locally inconsistent anomalies. Second, this loss is preceded by a measurable collapse in the effective rank of chunk embeddings. Third, a pseudo-label-free consistency objective on residual-preserving views avoids both failures. It also costs no more than uniform edge dropping, which is linear in the number of edges.
Research Program
The method has four pieces, each borrowed from one of the input directions and adapted to this setting.
Feature decomposition. The first piece comes from ABNet. ABNet projects a node's current features orthogonally to its historical states to expose abrupt behavioral change, and its ablation shows this projection matters: Wikipedia AUC falls from 92.21% to 71.46% without it. We reuse the projection to split each node's current features $x_v^t$ into a history-explained part $h_v^t$ and an orthogonal residual $r_v^t$. ABNet uses this split to build a detector and to synthesize anomalies by injecting noise into a latent feature space. We use it instead to decide what an augmentation is allowed to touch.
View construction. Views are built so that the residual is never touched. Feature masking applies only to the coordinates of $h_v^t$. Each edge $(u,v)$ in the chunk is dropped with probability $p_{uv} = p_{\max}\,\sigma(\mathrm{cos}(r_u^t, r_v^t))$, where $\sigma$ rescales cosine similarity to $[0,1]$. Edges between nodes whose recent deviations agree are "smooth" and safe to perturb. Edges that connect disagreeing nodes carry the local inconsistency and are protected.
This follows the logic of centrality-aware dropping in GCA, which protects hub edges based on degree or betweenness on a static graph, but with a different target. We protect the edges that carry inconsistency, not the structurally central ones, and the rule costs one similarity per edge, not a centrality computation. The construction also answers, for the adaptation setting, the tension raised by LiSA and by AC2L-GAD. LiSA shows unguided augmentation shifts labels and uses expensive bi-level generators to keep subgraphs label-invariant. AC2L-GAD deliberately generates counterfactual anomalies on static graphs. We need label invariance only, and we obtain it structurally, by never perturbing the signal that would change the label, without a generator.
Adaptation objective. Each arriving chunk $t$ yields two views of every node's temporal ego-graph (the node plus its recent multi-hop neighborhood). The detector $f_\theta$ outputs an embedding $z_v$ and an anomaly score $a_v$. We update only lightweight adapter and normalization parameters, minimizing
$$\mathcal{L}_t = \frac{1}{|V_t|}\sum_{v \in V_t}\Big[(a_v^{(1)} - a_v^{(2)})^2 + \lambda\,\lVert \bar z_v^{(1)} - \bar z_v^{(2)}\rVert^2\Big],$$
where superscripts index the two views, bars denote $\ell_2$-normalized embeddings, and $\lambda$ weights embedding agreement against score agreement. No pseudo-labels and no neighbor-agreement filter are used.
Collapse guard. A positive-pairs-only objective like this can collapse. The guard comes from Are we measuring oversmoothing in graph neural networks correctly?. That paper shows Dirichlet energy, a sum of feature differences across edges, misses collapse in trained networks. Effective rank, by contrast, tracks accuracy with an average Pearson correlation of 0.91 across depth. Effective rank here is the exponential of the entropy of the normalized singular values of the chunk embedding matrix $Z_t$. We accept a candidate update $\theta'$ only if
$$\mathrm{Erank}(Z_t^{\theta'}) \ge \rho\,\mathrm{Erank}(Z_t^{\theta_0}),$$
where $\theta_0$ is the source model and $\rho \in (0,1)$ is a tolerance. Otherwise the update is rolled back.
That paper validates rank as a depth diagnostic, not as an adaptation diagnostic, so part of our analysis tests whether the correlation transfers. It also warns that high rank does not imply good representations. For that reason the guard is only a veto and never an objective to maximize.
Backbones. The method is backbone-agnostic. We run it on a TADDY-style dynamic graph Transformer (TADDY) and on the OTTA-DGAD backbone, so that any difference from OTTA-DGAD comes from the adaptation rule rather than the encoder.
Baselines
The naive reference point is the frozen source detector: the same backbone trained on the source domain or period and applied unchanged to every target chunk. Any adaptation method must beat it. Otherwise adaptation is only adding risk.
The strongest existing competitor is OTTA-DGAD, the only method in the file that performs online test-time adaptation for dynamic graph anomaly detection. It reports consistent AUROC and AUPRC gains and lower cross-chunk variance than existing detectors and generalist models on ten datasets. We found no per-dataset reported result on Lacuna for OTTA-DGAD, so we will rerun it under its own "test-then-adapt" chunk protocol. It will use the same backbone, chunk boundaries, and source checkpoints as our method. Its buffer sizes will be tuned with the same number of trials we spend on $p_{\max}$, $\lambda$ and $\rho$, selected on a held-out source-side validation stream. Beating it means higher mean AUPRC across chunks with no higher variance, on a majority of the target streams. It also means a statistically clear gain on the high-inconsistency anomaly stratum described below.
For absolute calibration on shared benchmarks, the reported figures are:
- ABNet: 92.21% AUC on Wikipedia, 75.78% on Reddit and 75.34% on Mooc (ABNet).
- TGAT: 0.792 AUC on DGraph (DGraph).
- TADDY reports only relative gains (an average 4.49% AUC over the best baseline) and STCAD reports averages over TADDY. No absolute numbers for those two appear in the file.
We also run three ablations that isolate the key ingredients:
- Uniform views. Replace residual-preserving views with uniform edge dropping and feature masking at matched rates. This isolates the semantic-preserving choice.
- No guard. Remove the rank guard. This isolates collapse control.
- Neighbor-agreement filter restored. Add OTTA-DGAD's filter on top of our consistency loss. This isolates the harm we hypothesize.
Why not just…?
Why not just keep the source model frozen and recalibrate the score threshold per chunk? This is the cheapest option and a real threat if drift is mostly a score shift. But recalibration is a monotone transformation within a chunk, so it cannot change per-chunk AUROC or AUPRC at all. It can only help pooled metrics and threshold-based F1. We will report both per-chunk and pooled metrics. If our method gains only on pooled metrics, the recalibration baseline will expose that, and the claim reduces to calibration.
Why not just use OTTA-DGAD's pseudo-labeling, perhaps with entropy alone? OTTA-DGAD's own ablation shows entropy-only pseudo-labels hurt, because models are confidently wrong under shift. Its fix, requiring agreement with neighbors, is exactly the rule that should discard isolated anomalies. That conflict is the premise of GADAM's inconsistency mining. We test this directly. Target anomalies are binned by local inconsistency, measured as the distance of each anomaly's raw score from its neighbors' mean. If OTTA-DGAD's recall in the top bin is not lower than ours, our central hypothesis is wrong.
Why not just use standard uniform augmentation consistency? Uniform dropping is cheap and widely used. But AC2L-GAD argues that random perturbation erases the subtle anomalous signal, and LiSA documents label shift under unguided edits. The uniform-views ablation settles this at identical cost. We expect it to stay near the frozen baseline or below it on anomalies with small residuals, but it is plausible it works comparably on burst-type anomalies. We will report that honestly by anomaly type.
Why not just swap in an anti-smoothing backbone and skip adaptation? Architectures such as A-DGN, which keeps deep networks stable through anti-symmetric weight updates, or FB-GNN, which adds a high-pass channel to ordinary neighbor averaging, prevent smoothing across layers. They do nothing about drift, and they were evaluated only on static graphs. We include a frozen FB-GNN-augmented backbone as an extra reference. If it matches adapted models under cross-domain shift, adaptation is unnecessary for these streams.
Proposed Experiments
Streams and shifts. We use four settings:
- OTTA-DGAD's cross-domain chunked protocol over its ten-dataset suite.
- Cross-dataset shifts among TADDY's benchmarks (UCI Message, Digg, Bitcoin-Alpha, Bitcoin-OTC, Email-DNC, AS-Topology) at 1%, 5% and 10% injected anomalies.
- Wikipedia, Reddit and Mooc with organic labels, following ABNet.
- A temporal-drift stream built from DGraph by chunking its timestamped edges. DGraph has 3.7M nodes and a 1.3% anomaly rate, and serves as the scalability test.
Metrics. We report per-chunk AUROC and AUPRC (mean and variance), pooled metrics, stratified recall by local inconsistency, wall-clock time per chunk and peak memory.
Decisive checks. Three results decide whether the proposal holds.
- Diagnostic. Under OTTA-DGAD, a drop in effective rank should precede drops in AUPRC by at least one chunk, while Dirichlet energy stays flat. If rank does not lead AUPRC, the guard has no grounding and is dropped.
- Stratum test. The high-inconsistency recall gap described above must appear.
- Joint ablation. Uniform views must underperform residual-preserving views specifically on anomalies whose residual norm is small. That is where erasing the residual should matter most.
If all three hold, and the method matches uniform dropping in cost on DGraph, the claim stands that adaptation must be designed around what it is forbidden to smooth.
References
- Online Test-Time Adaptation for Generalizable Dynamic Graph Anomaly Detection
- Boosting Graph Anomaly Detection with Adaptive Message Passing
- ABNet: Mitigating Sample Imbalance in Anomaly Detection Within Dynamic Graphs
- Graph Contrastive Learning with Adaptive Augmentation
- Mind the Label Shift of Augmentation-based Graph OOD Generalization
- AC2L-GAD: Active Counterfactual Contrastive Learning for Graph Anomaly Detection
- Are we measuring oversmoothing in graph neural networks correctly?
- Anomaly Detection in Dynamic Graphs via Transformer
- DGraph: A Large-Scale Financial Dataset for Graph Anomaly Detection
- Anti-Symmetric DGN: a stable architecture for Deep Graph Networks
- Complete the Missing Half: Augmenting Aggregation Filtering with Diversification for Graph Convolutional Networks