acceptodds
Under review as a conference paper at ICLR 2027

RAFlow: Graph-Conditioned Continuous Normalizing Flows for Relational Time-Series Anomaly Detection

Abstract

Multivariate time-series anomaly detection requires modelling both temporal dynamics and evolving dependencies among correlated entities. Existing graph-based normalizing flow methods incorporate learned relational structures into density estimation, but condition the graph only before the flow transformation, limiting the ability of the likelihood model to adapt continuously to dynamic relational structure. We propose RAFlow, a graph-conditioned continuous normalizing flow framework for relational time-series anomaly detection, where the velocity field of a neural ordinary differential equation (ODE) is jointly conditioned on learned temporal and relational embeddings at every integration step, ensuring inter-entity dependencies continuously shape the density transformation throughout the entire trajectory. Terminal likelihood alone cannot distinguish anomalies that reach similar density regions through irregular latent trajectories; we introduce a path-integral anomaly score that combines terminal log-likelihood with the kinetic energy of the ODE trajectory as complementary detection signals. Theoretically, we prove that RAFlow yields a consistent anomaly score that is stable, robust to measurement noise, and resistant to wild fluctuations under small input variations. Extensive experiments on five benchmark datasets against nineteen baselines show that RAFlow consistently outperforms existing graph-based flow methods and achieves the strongest performance on densely correlated benchmarks, highlighting the effectiveness of continuous graph-conditioned density estimation.

Primary area:
learning on time series and dynamical systems
Keywords:
Time-Series Anomaly Detection, Continuous Normalizing Flows, Neural ODEs, Graph Neural Networks, Dynamic Graph Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_CHHKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Relational Adaptive Flow (RAFlow) is a multivariate time-series anomaly detection method, built upon graph-conditioned continuous normalizing flows. It introduces window-adaptive relational context, entity-specific Gaussian priors, kinetic-energy regularization, and path-integral anomaly scoring.

Critical strengths
  1. Reframing relational anomaly detection as assessing both conditional density and trajectory energy, rather than relying on likelihood alone, is fancy.
  2. Using AUC-PR and VUS-PR makes the evaluation more informative than relying on point-adjusted detection metrics.
Critical weaknesses
  1. The discrete-flow baseline uses substantially fewer parameters than RAFlow, so the headline gain may reflect increased model capacity rather than better flow dynamics.

  2. The obvious alternative of conditioning every discrete-flow layer on the same relational context is insufficiently specified, leaving the core premise unresolved.

  3. Ablation only partly validates the paper’s story, leaving unclear whether the training regularizer, not the trajectory-energy score, drives most of the performance.

  4. The method’s central stability guarantee presupposes that Gaussian log-density is globally one-Lipschitz, so it fails even for identity flows and remains invalid in exactly the simple cases that its assumptions explicitly allow.

  5. Graph & Discrete Flow comparison relies on baselines from 2020–2023, leaving it unclear whether the reported gains hold against more recent methods.

  6. Case studies mainly illustrate smoother scores rather than clearer anomaly separation and raise concerns about possible false positives.

Questions for authors

Q1. Could you clarify how the conditional consistency argument is justified when anomalous windows need not remain separated after context construction?

Q2. I’m curious whether trajectory energy gives anything beyond ordinary regularization in this setting. It is presented as an additional anomaly signal, but the reported ablation does not separate its training and scoring roles.

6 comments
AuthorsOpenReview ↗

Discrete-Flow Controls, Stability, and Anomaly Scoring

Thank you for the detailed review. We corrected the parameter count, clarified the discrete controls, and revised the theoretical guarantees.

W1–W2. Capacity and conditioning every discrete layer

The SMD checkpoint has 46,569 parameters (0.047M). The original efficiency table overstated this count. Standard MTGFlow has a reported 4.1M parameters. The four residual controls and MTGFlow-small each use about 0.05M.

D-RF(MM) uses M=2,4,8M=2,4,8 untied residual layers, each conditioned on the same window context CC. Each layer adds gθm(z,m/M,C)/Mg_{\theta_m}(z,m/M,C)/M to the latent state and is invertible for fixed CC when its residual branch is contractive. Tied Euler(2) shares one field across two updates. RAFlow trains with adaptive ODE integration and uses two tied Euler updates at inference.

Five-seed mean test AUC-PR (%).

ModelSMDSMAPMSLPSMSWaT
MTGFlow-small54.872.475.268.078.4
D-RF(2)58.074.678.071.280.8
D-RF(4)59.075.278.872.081.4
D-RF(8)59.275.479.072.281.4
Tied Euler(2)58.475.078.271.481.0
RAFlow61.477.281.474.683.2

RAFlow leads MTGFlow-small by 4.8–6.6 points and D-RF(8) by 1.8–2.4 points. D-RF(8) uses the same context at every layer and has a similar parameter count. The comparison still varies weight sharing and training integration.

W3 / Q2. Kinetic energy during training and inference

In Table 13, we train with the kinetic penalty off or on and score each checkpoint with composite NLL alone and with weighted trajectory energy added. The first four rows give five-seed mean test AUC-PR (%). Contrasts are in percentage points.

Training / scoreSMDSMAPMSLPSMSWaT
Off / NLL60.074.478.472.882.0
Off / NLL + energy59.875.279.272.281.8
On / NLL61.277.081.274.483.0
On / NLL + energy61.477.281.474.683.2
Training main effect+1.4+2.3+2.5+2.0+1.2
Scoring main effect0.0+0.5+0.5−0.20.0
Training × scoring+0.4−0.6−0.6+0.8+0.4

We average each main effect across the levels of the other factor. At the tested weights, the training effect is +1.2 to +2.5 points and the scoring effect ranges from −0.2 to +0.5 points. Weighted energy has 6–8% of the NLL score's standard deviation and a Spearman correlation of 0.66–0.74 with NLL. Comparison with a tuned conventional regulariser remains open.

W4. Gaussian NLL and stability

The identity-flow counterexample is correct. The standard Gaussian NLL has unbounded gradient xx. Proposition 1 bounds the score on bounded inputs under regularity conditions on the velocity, its divergence, and the context map. The following comments provide the revised theorem statements, proofs, corollaries, and supporting lemmas.

W5. Recent baselines

MTGFlow-cluster (Zhou et al., IEEE TKDE 2024) scores 59.0 / 76.2 / 79.4 / 72.6 / 82.2 test AUC-PR (%) in dataset order. RAFlow leads by 1.0–2.4 points. Appendix K.7 also reports GraphTSAD at 58.4 / 75.8 / 78.8 / 71.4 / 81.8. SARAD and CATCH lead the broader benchmark on SWaT and SMAP, respectively.

W6. Test-set false alarms at matched recall

DatasetMethodFPR at 80% recallFPR at 90% recallEpisodes
SMDRAFlow8.8%14.4%13.0
SMDMTGFlow11.4%18.2%19.8
SMDGANF14.2%22.4%26.6
MSLRAFlow7.2%12.6%10.6
MSLMTGFlow10.2%17.4%18.2
MSLGANF12.8%21.0%25.4

Episodes are counted per 1,000 normal windows at 80% recall. RAFlow yields 34% fewer episodes on SMD and 42% fewer on MSL than MTGFlow at this recall.

Q1. Consistency after context construction

Theorem 1 assumes a population margin in the final score, which depends on the window and its context. Under the stated assumptions, the fitted score gap is at least one-third of that margin with probability tending to one. On the test split, 20% of anomalous SMD contexts and 17% of anomalous MSL contexts fall within the 95% normal-context ellipsoid. On the validation split, the 5th percentile of anomalous scores exceeds the 95th percentile of normal scores by 0.34 on SMD and 0.40 on MSL. These quantile gaps describe the observed validation distributions.

Other corrections

  • Table 2. The discrete-flow ablation reduces AUC-PR by 3.4–7.4 points, with 3.4 on SMAP.
  • Appendix J. RAFlow takes about 1.6× MTGFlow's training time per epoch and 1.7× its inference latency per window.
  • Correlation. We report the dataset-level results and withdraw the claim that the advantage over MTGFlow scales with correlation strength.
  • ODE analysis, panel c. The value of about 0.85 is an empirical Lipschitz estimate from power iteration.
  • Section 3.4. Higher expected trajectory energy for anomalous windows is conditional on the displacement gap in Theorem 2.
AuthorsOpenReview ↗

Response to Reviewer CHHK — Theorem 1 and Proof Steps (i)–(v)

Notation. KK is the number of entities,

\phi_\theta=\sum_k\phi_{k,\theta},$$ and $$\widetilde{\mathcal L}_N(\theta)=\frac{1}{NK}\sum_{c=1}^N\phi_\theta(X^c).$$ $\mathcal D\subset\mathbb R^{K\times T}$ is a compact set containing the supports of $P_N$ and $P_A$. **Theorem 1 (conditional anomaly score consistency)** **Regularity and mixing** (a) $\Theta$ is compact, $P_N,P_A$ are supported in $\mathcal D$, and $(\theta,X)\mapsto\phi_\theta(X)$ is continuous on $\Theta\times\mathcal D$. (b) The normal windows $X^1,\dots,X^N$ are strictly stationary with $\alpha$-mixing coefficients $\alpha(h)$ satisfying $\sum_{h\ge1}\alpha(h) < \infty$. **Identifiable minimiser** $\mathcal R(\theta)=\frac{1}{K}\mathbb E_{P_N}[\phi_\theta(X)]$ has a unique minimiser $\theta^\ast$ with $$\inf_{\\|\theta-\theta^\ast\\|\ge\epsilon}\mathcal R(\theta) > \mathcal R(\theta^\ast)$$ for all $\epsilon > 0$. **Score separation** There are $\tau\in\mathbb R$, $m > 0$ with $$P_N(\phi_{\theta^\ast}\le\tau)=1 \quad\text{and}\quad P_A(\phi_{\theta^\ast}\ge\tau+m)=1.$$ **Theorem** Under the three assumptions above (and the Lipschitz Assumption of the paper, which makes the flow well defined), let $$\widehat\theta_N\in\arg\min_\Theta\widetilde{\mathcal L}_N.$$ Then (i) $$\sup_{X\in\mathcal D}|\phi_{\widehat\theta_N}(X)-\phi_{\theta^\ast}(X)|\to0$$ in probability, and (ii) for $X_A\sim P_A$, $X_N\sim P_N$: $$\mathbb P\bigl(\phi_{\widehat\theta_N}(X_A)\ge\phi_{\widehat\theta_N}(X_N)+m/3\bigr)\to1.$$ **Proof** Put $\widehat{\mathcal R}_N=\widetilde{\mathcal L}_N$. A minimiser exists because $\widehat{\mathcal R}_N$ is continuous on the compact set $\Theta$ (we take a measurable selection). **(i).** Since $\phi$ is continuous on the compact set $\Theta\times\mathcal D$, hence bounded, $|\phi|\le M_\phi$, and uniformly continuous. Thus $$\omega(r):=\sup\\{ |\phi_\theta(X)-\phi_{\theta'}(X)| : X\in\mathcal D,\\ \\|\theta-\theta'\\|\le r \\}$$ is nondecreasing with $\omega(r)\to0$ as $r\downarrow0$. **(ii).** Fix $\theta$ and let $Y_c=\phi_\theta(X^c)/K$, so $|Y_c|\le M_\phi/K$ and $\mathbb EY_c=\mathcal R(\theta)$. For bounded $\alpha$-mixing variables, $$|\mathrm{Cov}(Y_c,Y_{c+h})|\le4(M_\phi/K)^2\alpha(h)$$ (Ibragimov's covariance inequality). Hence $$\begin{aligned} \mathrm{Var}\Bigl(\tfrac{1}{N}\sum_cY_c\Bigr) &\le\frac{1}{N^2}\Bigl[N\tfrac{M_\phi^2}{K^2}+2\sum_{h=1}^{N-1}(N-h)\\,4\tfrac{M_\phi^2}{K^2}\alpha(h)\Bigr]\\\\ &\le\frac{M_\phi^2}{NK^2}\Bigl(1+8\sum_{h\ge1}\alpha(h)\Bigr)\xrightarrow[N\to\infty]{}0, \end{aligned}$$ so $\widehat{\mathcal R}_N(\theta)\to\mathcal R(\theta)$ in probability by Chebyshev. **(iii).** Given $\epsilon > 0$ choose $r$ with $\omega(r)/K < \epsilon/3$ and a finite $r$-net $\theta_1,\dots,\theta_J$ of $\Theta$ (compactness). For any $\theta$ pick $j$ with $\\|\theta-\theta_j\\|\le r$. Then $$|\widehat{\mathcal R}_N(\theta)-\widehat{\mathcal R}_N(\theta_j)|\le\omega(r)/K < \epsilon/3,$$ and likewise for $\mathcal R$. Therefore $$\sup_\theta|\widehat{\mathcal R}_N-\mathcal R|\le\max_{j\le J}|\widehat{\mathcal R}_N(\theta_j)-\mathcal R(\theta_j)|+\tfrac{2\epsilon}{3},$$ and the maximum over finitely many terms tends to $0$ in probability (Step 2). So $$\mathbb P\bigl(\sup_\theta|\widehat{\mathcal R}_N-\mathcal R| < \epsilon\bigr)\to1$$ for every $\epsilon > 0$. **(iv).** Fix $\epsilon > 0$, let $\eta_\epsilon=\inf_{\\|\theta-\theta^\ast\\|\ge\epsilon}\mathcal R(\theta)-\mathcal R(\theta^\ast) > 0$ (Assumption Identifiable minimiser). On the event $\\{\sup_\theta|\widehat{\mathcal R}_N-\mathcal R| < \eta_\epsilon/2\\}$, for every $\theta$ with $\\|\theta-\theta^\ast\\|\ge\epsilon$, $$\widehat{\mathcal R}_N(\theta) > \mathcal R(\theta)-\eta_\epsilon/2\ge\mathcal R(\theta^\ast)+\eta_\epsilon/2 > \widehat{\mathcal R}_N(\theta^\ast),$$ so $\theta$ is not a minimiser. Hence $\mathbb P(\\|\widehat\theta_N-\theta^\ast\\| < \epsilon)\to1$. **(v).** For $X\in\mathcal D$, $|\phi_{\widehat\theta_N}(X)-\phi_{\theta^\ast}(X)|\le\omega(\\|\widehat\theta_N-\theta^\ast\\|)$. Given $\epsilon > 0$ choose $r$ with $\omega(r) < \epsilon$; on $\\{\\|\widehat\theta_N-\theta^\ast\\|\le r\\}$ (probability $\to1$ by Step 4) the supremum over $\mathcal D$ is $< \epsilon$.
AuthorsOpenReview ↗

Response to Reviewer CHHK — Theorem 1, Final Proof Step, and Corollary 1

Theorem 1, proof continued

(vi). Let EN=sup⁡D∣ϕθ^N−ϕθ∗∣≤m/3\mathcal E_N=\\{\sup_{\mathcal D}|\phi_{\widehat\theta_N}-\phi_{\theta^\ast}|\le m/3\\}; P(EN)→1\mathbb P(\mathcal E_N)\to1. By Assumption Score separation, almost surely ϕθ∗(XN)≤τ\phi_{\theta^\ast}(X_N)\le\tau and ϕθ∗(XA)≥τ+m\phi_{\theta^\ast}(X_A)\ge\tau+m with both points in D\mathcal D. On EN\mathcal E_N, ϕθ^N(XA)≥τ+m−m3\phi_{\widehat\theta_N}(X_A)\ge\tau+m-\tfrac m3 and ϕθ^N(XN)≤τ+m3\phi_{\widehat\theta_N}(X_N)\le\tau+\tfrac m3, so

ϕθ^N(XA)−ϕθ^N(XN)≥m−2m3=m3.\phi_{\widehat\theta_N}(X_A)-\phi_{\widehat\theta_N}(X_N)\ge m-\tfrac{2m}3=\tfrac m3.

Therefore the probability in (ii) is at least P(EN)→1\mathbb P(\mathcal E_N)\to1. (The event EN\mathcal E_N does not involve the test points, so no independence between the training sample and the test pair is needed.)

Why support separation alone is not enough

Let PNP_N and PAP_A have disjoint supports at distance ϵ0>0\epsilon_0 > 0. Since ϕθ∗\phi_{\theta^\ast} is Lipschitz on D\mathcal D with constant LϕL_\phi (see Proposition 1), points of suppPA\mathrm{supp}P_A within ϵ0\epsilon_0 of suppPN\mathrm{supp}P_N can differ from normal scores by at most Lϕϵ0L_\phi\epsilon_0, so m≤Lϕϵ0m\le L_\phi\epsilon_0 whenever Assumption Score separation holds, and no separation can hold for arbitrarily small ϵ0\epsilon_0. Conversely, even with a large distance, a trained density can place high mass between the clusters; then anomalies there score lower than normal points, and Assumption Score separation fails. Separation of the scalar score is therefore the quantity that matters.

Corollary 1 (entity-level separability)

Entity-level separation

For an entity kk there are τk\tau_k, mk>0m_k > 0 with PN(ϕk,θ∗≤τk)=1P_N(\phi_{k,\theta^\ast}\le\tau_k)=1, PA(ϕk,θ∗≥τk+mk)=1P_A(\phi_{k,\theta^\ast}\ge\tau_k+m_k)=1, and (θ,X)↦ϕk,θ(X)(\theta,X)\mapsto\phi_{k,\theta}(X) is continuous on Θ×D\Theta\times\mathcal D.

Corollary

Under Assumptions Regularity and mixing, Identifiable minimiser, Entity-level separation:

P(ϕk,θ^N(XA)≥ϕk,θ^N(XN)+mk/3)→1.\mathbb P\bigl(\phi_{k,\widehat\theta_N}(X_A)\ge\phi_{k,\widehat\theta_N}(X_N)+m_k/3\bigr)\to1.

Proof

Steps (i)–(iv) of the proof of Theorem 1 use only the training objective and give θ^N→θ∗\widehat\theta_N\to\theta^\ast in probability. Define ωk(r)\omega_k(r) as in Step (i) for ϕk\phi_k; it tends to 00 by uniform continuity on Θ×D\Theta\times\mathcal D. As in Step (v),

sup⁡D∣ϕk,θ^N−ϕk,θ∗∣≤ωk(∣θ^N−θ∗∣)→0\sup_{\mathcal D}|\phi_{k,\widehat\theta_N}-\phi_{k,\theta^\ast}|\le\omega_k(\\|\widehat\theta_N-\theta^\ast\\|)\to0

in probability. Step (vi) with (ϕk,τk,mk)(\phi_k,\tau_k,m_k) in place of (ϕ,τ,m)(\phi,\tau,m) gives the claim.

Detection, not localisation

The corollary concerns entity kk separately and says nothing about whether kk is the origin of the anomaly. Because Ck(X)C_k(X) depends on every entity, perturbing XjX_j (j≠kj\neq k) changes CkC_k and can change ϕk\phi_{k} with XkX_k fixed. Thus ϕk large\\{\phi_k\text{ large}\\} does not imply k is the source\\{k\text{ is the source}\\}, and the entity scores should be used as a ranking of entities to inspect.

Reviewer_LG8pOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RAFlow, a graph-conditioned continuous normalizing flow framework for multivariate time-series anomaly detection. The method combines LSTM-based temporal representations with learned relational graph embeddings to condition the velocity field of a neural ODE throughout integration. In addition to terminal negative log-likelihood, RAFlow incorporates trajectory kinetic energy into its anomaly score to capture irregular latent transformation dynamics. The paper also presents theoretical results concerning score consistency, robustness, and stability. Experiments on five benchmark datasets against nineteen baselines, together with ablation studies and efficiency analyses, demonstrate competitive anomaly detection performance, particularly on datasets characterized by strong inter-variable dependencies.

Critical strengths
  1. A meaningful integration of relational modeling and continuous density estimation.

The paper develops a graph-conditioned continuous normalizing flow in which temporal and relational representations directly influence the neural ODE velocity field. This provides a reasonable approach to incorporating inter-variable dependencies into density estimation beyond conventional graph-based conditioning before a flow transformation. The integration is technically meaningful for multivariate time-series anomaly detection, where temporal patterns and relationships between variables jointly determine normal system behavior.

  1. An interesting trajectory-aware anomaly score and relatively broad empirical evaluation.

The proposed anomaly score combines terminal negative log-likelihood with ODE trajectory kinetic energy, extending conventional density-based anomaly detection by considering the transformation path. This is an interesting direction because terminal likelihood alone may not fully characterize irregular transformation dynamics. The evaluation spans five benchmark datasets, nineteen baselines, and multiple performance metrics, including AUC-PR and VUS-PR. The inclusion of component ablations, hyperparameter analyses, and computational comparisons further strengthens the empirical scope of the paper.

Critical weaknesses
  1. The theoretical guarantees and the distinctive contribution of continuous graph conditioning require stronger justification.

The paper makes several theoretical claims regarding anomaly-score consistency, robustness, and stability. However, the stated assumptions do not appear sufficient to establish all the claimed conclusions. In particular, the joint support separation discussed in Theorem 1 does not by itself guarantee separation under a particular scalar anomaly score. The energy ordering in Theorem 2 also requires further justification, while the global stability claim in Proposition 1 appears to depend on additional regularity assumptions governing the flow transformation and likelihood function. These concerns are consequential because theoretical guarantees are presented as a central contribution of the framework.

The methodological distinction between continuous graph conditioning and continuously evolving relational structure also requires clarification. The graph context in Eq. (3) appears to be computed once per input window and held fixed during ODE integration in Eq. (4). Although this still permits the velocity field to depend on relational information at every integration step, it does not necessarily represent an evolving graph within the ODE trajectory. Furthermore, the numerical results indicate that relatively coarse Euler integration achieves performance close to more sophisticated solvers. Additional controlled comparisons would help establish whether the improvements arise specifically from the continuous formulation or from the underlying graph-conditioned transformation architecture.

  1. The experimental protocol and ablations do not sufficiently isolate the sources of performance improvements.

Appendix C reportedly uses validation AUC-PR for early stopping. Since AUC-PR requires anomaly labels, this introduces supervised information into model selection despite the paper's emphasis on unsupervised anomaly detection. While this does not necessarily invalidate unsupervised parameter training, the distinction between label-free training and label-assisted model selection should be made explicit. It is also important to establish whether all compared methods receive equivalent access to labeled validation information.

Moreover, the computational comparison indicates that RAFlow has approximately twice as many parameters as MTGFlow (8.3M versus 4.1M). Without capacity-matched comparisons, it is difficult to determine how much of the observed improvement should be attributed to continuous graph-conditioned modeling rather than increased model capacity.

The trajectory-aware scoring ablation presents an additional attribution problem. Removing kinetic energy appears to affect both training regularization and inference-time anomaly scoring. Consequently, the observed performance differences do not independently establish the contribution of energy-based scoring. Factorial ablations separating these two components would provide stronger evidence for the proposed detection objective.

Together, these issues limit confidence that the reported improvements can be attributed specifically to the paper's central methodological contributions.

Questions for authors

Q1. Theoretical guarantees: Could the authors clarify the assumptions required for Theorems 1 and 2 and Proposition 1? In particular, how does joint support separation guarantee consistency of the specific scalar anomaly score? What assumptions establish the claimed kinetic-energy ordering and global Lipschitz stability of the likelihood component? More explicit proofs or appropriately qualified conclusions would materially affect my assessment of the theoretical contribution.

Q2. Validation labels and model selection: Appendix C indicates that validation AUC-PR is used for early stopping. Does this procedure use ground-truth anomaly labels? If so, is the same label-assisted model selection protocol applied to all baselines? Can the authors provide results using a label-free selection criterion to demonstrate performance under a strictly unsupervised evaluation protocol?

Q3. Contribution of continuous graph conditioning: Since the graph representation appears fixed during each ODE integration, what specific capability does the proposed continuous conditioning provide beyond existing graph-conditioned flow architectures? Can the authors provide comparisons with discrete graph-conditioned flows under matched parameter counts and comparable computational budgets? The numerical results with coarse Euler integration make this distinction particularly important.

Q4. Independent contribution of trajectory-aware scoring: Can the authors separately ablate kinetic-energy regularization during training and kinetic-energy scoring during inference? A comparison of likelihood-only training and scoring, energy-regularized training with likelihood-only scoring, and the complete method would help establish whether the reported improvements arise from the training regularizer, the inference score, or their combination.

2 comments
AuthorsOpenReview ↗

Theoretical Clarifications and Additional Experiments

Thank you for the constructive review. We revised the theory and added checkpoint-selection, flow-control, and kinetic-energy experiments.

Q1. Theoretical guarantees

  • Theorem 1. Disjoint input supports alone do not imply separation by the composite score. We assume a population margin m>0m>0 in the final scalar score after context construction. Under compactness, score continuity, stationary normal windows with summable α\alpha-mixing coefficients, and a unique, well-separated population minimiser, the score fitted by minimising the unregularised composite risk converges uniformly in probability. When the uniform error is at most m/3m/3, the fitted anomaly–normal score gap is at least m/3m/3.
  • Theorem 2. Let D=∥z(1)−X∥22D=\lVert z(1)-X\rVert_2^2 and δN=EN[E−D]≥0\delta_N=\mathbb E_N[\mathcal E-D]\geq 0. Cauchy–Schwarz gives E≥D\mathcal E\geq D on the unit interval. If EA[D]−EN[D]>δN\mathbb E_A[D]-\mathbb E_N[D]>\delta_N, then EA[E]≥EA[D]>EN[E]\mathbb E_A[\mathcal E]\geq\mathbb E_A[D]>\mathbb E_N[\mathcal E]. When this condition holds for every entity, Corollary 2 shows that adding energy widens the expected anomaly–normal score gap.
  • Proposition 1. We replace the global stability claim with a bounded-domain result. Bounds BB, GG, and MM on the input, velocity, and prior-mean norms bound the unit-covariance Gaussian NLL gradient by B+G+MB+G+M on the reachable region. Lipschitz regularity of the velocity, its divergence, and the context map then bounds score perturbations.

Q2. Validation labels and checkpoint selection

Yes. RAFlow used labelled validation AUC-PR for early stopping and unlabelled normal windows for parameter updates. The original baselines used their own stopping rules. In the follow-up comparison, U selects checkpoints by loss on held-out normal validation windows, and S selects by labelled validation AUC-PR. Entries show test AUC-PR (%) as U / S.

MethodSMDSMAPMSLPSMSWaT
RAFlow60.8 / 61.476.4 / 77.280.6 / 81.474.0 / 74.682.6 / 83.2
MTGFlow57.8 / 58.274.8 / 75.678.0 / 78.871.0 / 71.881.2 / 81.6
GANF55.6 / 56.870.6 / 72.472.8 / 74.667.4 / 69.077.0 / 78.4
SARAD57.4 / 58.673.2 / 74.875.6 / 77.467.8 / 69.483.4 / 84.6
CATCH56.6 / 57.876.8 / 78.276.8 / 78.469.0 / 70.680.8 / 82.2

RAFlow's scores under U are 0.6–0.8 percentage points lower than under S. Under U, RAFlow leads on SMD, MSL, and PSM among the five methods and remains 1.4–3.0 points ahead of MTGFlow.

Q3. Graph conditioning, capacity, and compute

We fix the graph and context per window during the ODE solve. The CNF uses a shared context-conditioned velocity field and adaptive training integration. Inference uses two weight-tied Euler updates. On SMD, higher-order solvers added at most 0.4 AUC-PR points at 2–5× the inference cost.

Table 9 overstated RAFlow's parameter count. The SMD model has 46,569 parameters (0.047M), about 1.1% of MTGFlow's reported 4.1M. We also compare against a reduced MTGFlow and residual controls that receive the same context at every layer. Entries are five-seed mean test AUC-PR (%).

ModelParamsSMDSMAPMSLPSMSWaT
MTGFlow-small≈0.05M54.872.475.268.078.4
Tied Euler(2)≈0.05M58.475.078.271.481.0
D-RF(8), untied≈0.05M59.275.479.072.281.4
RAFlow0.047M61.477.281.474.683.2

RAFlow leads MTGFlow-small by 4.8–6.6 points and D-RF(8) by 1.8–2.4 points. The controls approximately match the parameter budget. Training integration and weight sharing vary. Timings cover the standard models only.

Q4. Kinetic energy in training and scoring

In Table 13, we train with regularisation off or on and score each checkpoint with composite NLL alone and with weighted trajectory energy added. The first four rows give five-seed mean test AUC-PR (%). Contrasts are in percentage points.

Training / scoringSMDSMAPMSLPSMSWaT
Off / NLL60.074.478.472.882.0
Off / NLL + energy59.875.279.272.281.8
On / NLL61.277.081.274.483.0
On / NLL + energy61.477.281.474.683.2
Training main effect+1.4+2.3+2.5+2.0+1.2
Scoring main effect0.0+0.5+0.5−0.20.0
Training × scoring+0.4−0.6−0.6+0.8+0.4

We average each main effect over the levels of the other factor. The interaction measures how training regularisation changes the scoring effect. At the tested weights, the training main effect is positive on all five benchmarks (+1.2 to +2.5 points), while the scoring main effect is smaller and varies by dataset (−0.2 to +0.5 points).

Thank you for your rebuttal. After careful consideration of the manuscript quality, I have decided to retain my original score.

Reviewer_4AhxOpenReview ↗
Rating
1(−1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper uses continuous normalizing flow to address the problem of multivariate time series anomaly detection. Authors propose RAFlow, which employs a dynamically learned graph structure to describe the relationships among sensors. Time features are then fused with graph relationships to form a context, which is used to construct a velocity field via CNF. Kinetic energy from ODE trajectories is also used to construct an anomaly score. The main issues lie in the inconsistency between the contributions emphasized by the authors and the model implementation, as well as flaws in the theoretical proofs.

Critical strengths
  1. Anomalies in multi-sensor systems may indeed manifest as anomalies in the relationships between variables rather than as anomalies in individual values. This is a valid point.
  2. The proposed method is an end-to-end anomaly detection architecture.
  3. The experiments provide comprehensive coverage.
Critical weaknesses
  1. The dynamic graph does not appear to evolve continuously within the ODE. In Section 3.1, both AcA^c and CcC^c are fixed input conditions calculated before integration begins. So the edge weights are not recalculated during the ODE process, nor is the graph updated based on the latent state. This contradicts the authors’ claim of 'continuous evolution'.
  2. There is an inconsistency in the authors’ statements regarding the conditional likelihood. In Eq. (2) and elsewhere, the authors acknowledge that the training objective is the Composite Conditional Scoring Function, not the Fully Coupled Joint Density. However, in Appendix B.3, they state that under the Conditional Independence Assumption, the optimization objective is the joint likelihood function.
  3. Corollary 1 is not sufficiently proven. The authors claim that the per-entity score aids in locating the source of anomalies. However, which sensor is the source of the anomaly cannot be determined based solely on the individual conditional NLL. Even if a sensor has a very high anomaly score, it does not necessarily mean that it is the original fault node causing the anomaly.
  4. The proof of Theorem 2 in Appendix A appears to have some issues. The authors derive EPA[E]>EPN[E]E_{P_A}[E] > E_{P_N}[E] from EPA[E]>EPN[E]−δNE_{P_A}[E] > E_{P_N}[E] - \delta_N, but this step is not valid. the derivation process needs to be checked.
Questions for authors

Please refer to Weaknesses.

3 comments
AuthorsOpenReview ↗

what would be title? Response to Reviewer 4Ahx: Graph Conditioning and Theoretical Corrections

Thank you for identifying where the theoretical claims and experimental comparisons needed clarification. We address each question below.

Q1. Theoretical guarantees

ResultAssumptions and conclusion
Theorem 1Disjoint window supports alone are insufficient. We assume a population margin m>0m>0 for the scalar score after context construction. Under compactness, score continuity, stationary normal windows with summable α\alpha-mixing coefficients, and a unique, well-separated population minimiser, minimising the unregularised composite risk gives uniform score convergence in probability. An error of at most m/3m/3 on each score preserves a gap of at least m/3m/3.
Theorem 2Let D=∥z(1)−X∥22D=\lVert z(1)-X\rVert_2^2 and δN=EN[E−D]\delta_N=\mathbb E_N[\mathcal E-D]. If EA[D]−EN[D]>δN\mathbb E_A[D]-\mathbb E_N[D]>\delta_N, then EA[E]≥EA[D]>EN[D]+δN=EN[E]\mathbb E_A[\mathcal E]\geq\mathbb E_A[D]>\mathbb E_N[D]+\delta_N=\mathbb E_N[\mathcal E]. The first inequality follows from Cauchy–Schwarz on the unit interval.
Corollary 2If Theorem 2’s condition holds for every entity, adding energy widens the expected score gap by λ\lambda times the average energy gap.
Proposition 1We withdraw the global claim. With bounded inputs, velocity, and prior means, the Gaussian NLL gradient is bounded by B+G+MB+G+M on the reachable region. Lipschitz regularity of the velocity, divergence, and context map gives score stability on this bounded domain.

Q2. Validation labels and model selection

Yes. RAFlow used anomaly labels for early stopping on validation AUC-PR. Parameter updates used normal windows without anomaly labels. Baselines followed their original configurations, so the submitted comparison used different checkpoint-selection procedures. Independence of the labelled validation split from the test set remains unverified. The shared label-free comparison is pending. It would select checkpoints by validation loss on held-out normal training windows and reserve test labels for final evaluation.

Q3. Continuous conditioning, capacity, and compute

The graph context is computed once per window and held fixed during integration. The CNF shares one velocity field across depth, supports adaptive training integration, and defines a trajectory energy. Two-step Euler inference applies two weight-tied residual updates with the same context. On SMD, higher-order solvers added at most 0.4 AUC-PR points at 2–5× the inference cost. This supports two-step Euler as an inference choice.

The implementation audit gives 46,569 parameters at width 32 and 115,401 at width 64, calling the original 8.3M figure into question. The saved run configuration is needed to link an audited count to the SMD results. The original A100 timings for RAFlow and MTGFlow were 51.4 vs. 32.1 min/epoch and 8.1 vs. 4.8 ms/window. A matched comparison with separately trained discrete flows that receive the same context at every layer remains pending.

Q4. Training regularisation versus inference scoring

The reported ablation removed both kinetic-energy terms together.

AUC-PR (%, mean over five seeds)SMDSMAPMSLPSMSWaT
Full RAFlow61.477.281.474.683.2
Both energy terms removed60.074.478.472.882.0

The combined drop is 1.2–3.0 percentage points. The proposed 2×22\times2 ablation trains with the regulariser off and on, then scores each checkpoint with composite NLL alone and with NLL plus energy. Rescoring the same checkpoint isolates the inference effect. The individual effects and their interaction remain open.

Thank you for the clarification. However, I still got issues of Theorem 1. The authors acknowledge that the composite conditional score does not correspond to a valid joint density, yet still claim that joint support separation guarantees anomaly-score consistency. Then how does the separation of the joint distributions imply that the proposed composite score can distinguish normal and anomalous samples? Two distributions can have disjoint joint supports while sharing identical marginal distributions, and the stated assumptions do not establish that the learned conditional contexts preserve this distinction.

AuthorsOpenReview ↗

Follow-up to Reviewer 4Ahx — Theorem 1 and Empirical Score Separation

Thank you for following up and clarifying your concern. We have revised Theorem 1 to assume a margin in the final composite conditional score after context construction. As you point out, normal and anomalous windows can receive identical scores even when their joint supports are disjoint. The revised theorem shows that, if the population score has a margin m>0m>0, the fitted anomaly–normal score gap is at least m/3m/3 with probability tending to one.

Counterexample

Let normal windows be uniform on {(0,0),(1,1)}\{(0,0),(1,1)\} and anomalous windows be uniform on {(0,1),(1,0)}\{(0,1),(1,0)\}. The joint supports are disjoint, but each coordinate has the same Bernoulli(1/2)(1/2) marginal in both groups. If the context is constant and the fitted conditionals recover the normal marginals, the composite score for every window is

−log⁡(1/2)−log⁡(1/2)=2log⁡2.-\log(1/2)-\log(1/2)=2\log 2.

The score ties all four windows. In this example, the constant context leaves the score dependent only on the marginal distributions, so it cannot distinguish the two dependence patterns.

Revised Theorem 1

The proof treats the composite conditional negative log-likelihood as a scalar scoring function. Let the population score after context construction be

ϕ∗(X)=−∑k=1Klog⁡pθ∗(Xk∣Ck(X)).\phi^\ast(X)=-\sum_{k=1}^{K}\log p_{\theta^\ast}(X_k\mid C_k(X)).

For some threshold τ\tau and margin m>0m>0, the revised theorem assumes

P_A\bigl(\phi^\ast(X)\geq\tau+m\bigr)=1.$$ Under compactness, score continuity, stationary normal windows with summable mixing coefficients, and a unique, well-separated population minimiser, the score fitted by minimising the unregularised composite risk converges uniformly in probability to $\phi^\ast$. When the uniform error is at most $m/3$, the fitted score gap between an anomalous window $X_A$ and a normal window $X_N$ satisfies $$\widehat{\phi}_N(X_A)-\widehat{\phi}_N(X_N) \geq m-2(m/3)=m/3.$$ The probability of this uniform error bound tends to one. The result applies to the final score, which depends on both the observed window and its context, and requires only the stated score-margin assumption for separation. **Empirical diagnostics** Appendix K.6 reports two diagnostics. On the **test split**, we measure the fraction of anomalous contexts inside the 95% Mahalanobis ellipsoid of normal contexts. On the **validation split**, we compute the gap between the 5th percentile of anomalous composite scores and the 95th percentile of normal composite scores. | Dataset | Test anomalous contexts inside the normal-context ellipsoid | Validation score quantile gap | | :--- | ---: | ---: | | SMD | 20% | +0.34 | | MSL | 17% | +0.40 | Some anomalous test contexts occupy the normal-context region on both datasets. Separately, the validation composite scores show positive quantile gaps on both benchmarks. These gaps describe the observed validation distributions. Theorem 1 assumes an almost-sure margin in the population score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

RAFlow Revision Summary - Responses, New Experiments, and Revised Proofs

Dear Reviewers, Area Chair, and ICLR Program Chairs,

Thank you for the detailed feedback on RAFlow (submission 59838). We have uploaded the revised manuscript to OpenReview. This comment summarises the changes shared across the reviews. The reviewer-specific replies and follow-up proof comments provide the detailed responses, tables, and derivations.

1. Window-adaptive graph and discrete-flow controls

RAFlow computes the graph and context once per window and holds them fixed during ODE integration. The context conditions the velocity field at every integration step. We corrected the wording throughout the paper and clarified that two-step Euler inference is a two-layer, weight-tied graph-conditioned residual transformation. On SMD, higher-order solvers add at most 0.4 AUC-PR points at 2–5× the inference cost. We withdrew the claims that Euler isolates the effect of continuity and that RAFlow strictly generalises GANF and MTGFlow.

We added discrete residual flows with 2, 4, and 8 untied layers, each receiving the same context, and a tied two-step Euler control. We also tested RAFlow-J, which recomputes the graph from the latent state during integration. RAFlow-J has higher mean AUC-PR on PSM and SWaT, lower means on SMD, SMAP, and MSL, and higher variability on all five datasets.

2. Parameter accounting and capacity

The reported SMD checkpoint has 46,569 trainable parameters (0.047M). This corrects the original 8.3M entry. Standard MTGFlow has 4.1M parameters. At approximately 0.05M parameters, RAFlow leads MTGFlow-small by 4.8–6.6 AUC-PR points, the same-context D-RF(8) by 1.8–2.4 points, and tied Euler(2) by 2.2–3.2 points. Widening RAFlow to 0.44M adds 0.2 points per dataset at the reported precision. These comparisons control for parameter budget and, for the residual controls, access to context. Weight sharing and training integration still differ.

3. Label-free checkpoint selection

The submitted RAFlow runs used normal windows for parameter updates and labelled validation AUC-PR for checkpoint selection. The revised appendix applies common label-free and label-assisted selection rules to five methods. With label-free selection on held-out normal windows, RAFlow's mean test AUC-PR changes from 75.6% to 74.9%. It leads on SMD, MSL, and PSM among these methods and remains 1.4–3.0 points ahead of MTGFlow across all five datasets.

4. Kinetic energy, recent baselines, and false alarms

The five-seed 2×2 ablation trains with and without the kinetic penalty and evaluates each checkpoint with and without energy in the score. The training penalty adds 1.4–2.4 AUC-PR points, with the reported 95% confidence intervals above zero. Adding energy at inference gives +0.2 points, with all intervals spanning zero. The measured gain at the tested weights comes primarily from training regularisation.

The appendix adds recent graph- and flow-based comparisons, including MTGFlow-cluster. RAFlow leads this variant by 1.0–2.4 AUC-PR points. At 80% recall, its false-positive rates are 8.8% on SMD and 7.2% on MSL, compared with 11.4% and 10.2% for MTGFlow. It produces 34% and 42% fewer false-alarm episodes, respectively. On these datasets, 20% and 17% of anomalous contexts fall inside the normal-context ellipsoid. The empirical gaps between the 5th anomalous and 95th normal score percentiles are +0.34 and +0.40, respectively.

5. Revised claims and proofs

  • Theorem 1 assumes a margin directly in the final scalar score after context construction. Under the stated assumptions, estimation preserves at least one-third of that margin with probability tending to one. The proof uses the composite conditional risk and requires no joint-density interpretation. Disjoint joint supports alone do not guarantee score separation.
  • Theorem 2 and Corollary 2 give an explicit displacement-gap condition involving the excess kinetic energy of normal windows. Under this condition, the energy term widens the expected score gap.
  • Proposition 1 gives a Lipschitz stability bound on bounded inputs under explicit regularity assumptions on the velocity, its divergence, and the context map.
  • Corollary 1 concerns entity-level detection under a score-margin assumption. We describe entity scores as a diagnostic ranking and remove the claim that these scores identify the source of a fault.

The revised manuscript includes the statements and proofs, and Appendix K reports the new experiments. We also corrected the discrete-flow ablation range and runtime description, qualified the correlation claim, and identified the ODE Lipschitz value as a power-iteration estimate. Thank you for considering the revised submission.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.