acceptodds
Under review as a conference paper at ICLR 2027

RAGMO-NET: REGIME-AWARE GRAPH MEMORY OPTIMIZATION NETWORK FOR EEG

Abstract

Dynamic graph representation methods have shown promise for modeling electroencephalography (EEG) connectivity, but existing approaches are limited by heuristic memory updates, costly matrix-valued recurrent state transitions, and a lack of principled adaptation under non-stationary dynamics. We propose Regime-Aware Graph Memory Optimization Network (RAGMO-Net), a framework that adapts graph memory updates based on latent temporal regimes corresponding to evolving connectivity states. RAGMO-Net introduces a regime-conditioned affine graph-memory cell whose state transition is a scalar multiple of the identity, giving a constant-cost update per graph snapshot without matrix-valued gates or nonlinear state transitions. To provide a principled foundation for memory adaptation, we further formulate memory refinement as a regime-aware proximal refinement process, linking dynamic graph learning with non-stationary optimization. Experiments on both EEG and non-EEG datasets demonstrate consistent improvement over strong dynamic graph baselines. These results show that combining regime awareness with optimization-driven memory adaptation yields a scalable and robust approach for learning from non-stationary dynamic graphs.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Electroencephalography (EEG), Graph Memory, Brain Connectivity Network, Mirror Descent Optimization

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ehBtOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the challenges of inefficient memory updates and non-stationary dynamics in EEG-based graph learning. To address these issues, the authors propose RAGMO-Net, a regime-aware graph memory optimization framework. Specifically, it introduces a scalar-gated affine memory mechanism for efficient temporal modeling and a regime-aware proximal refinement strategy for stable memory adaptation. The authors further provide theoretical analyses of memory stability and convergence properties. Experimental results on EEG and non-EEG datasets demonstrate improved classification performance and computational efficiency compared with existing graph learning methods.

Critical strengths
  1. Meaningful research problem. The paper addresses the challenges of non-stationary graph dynamics and computationally expensive memory updates, which are relevant to EEG analysis and dynamic graph learning.

  2. Efficient memory architecture. The proposed scalar-gated affine memory mechanism simplifies temporal state transitions and reduces computational overhead compared with conventional recurrent architectures.

  3. Extensive experimental evaluation. Experiments across EEG and non-EEG datasets, together with ablation studies and runtime comparisons, demonstrate the effectiveness and computational efficiency of the proposed framework.

Critical weaknesses
  1. The scalar-gated affine update resembles conventional gated interpolation mechanisms. The authors should further clarify its methodological novelty compared with existing gated recurrent models and adaptive memory architectures.

  2. Although the model introduces a learned regime score, it remains unclear whether the detected regimes correspond to meaningful changes in EEG connectivity. Further quantitative validation of regime detection would strengthen the proposed framework.

  3. The stability and contraction results rely on bounded representations and additional Lipschitz-related assumptions. The authors should clarify whether these conditions hold in practice and discuss their implications for model behavior.

Questions for authors
  1. The analysis mainly focuses on regime scores and connectivity visualizations. Further investigation of their relationships with established EEG biomarkers would strengthen the neurological interpretation of the learned dynamics.

  2. Although subject-wise cross-validation is employed, additional cross-session or cross-dataset EEG evaluations would provide stronger evidence of robustness to non-stationary neural dynamics.

  3. Figure 2 is overly simplistic and does not adequately illustrate the proposed methodology.

  4. Most of the dynamic graph learning methods used for comparison were proposed more than three years ago. The authors should include more recent state-of-the-art methods as baselines to strengthen the experimental evaluation.

3 comments
AuthorsOpenReview ↗

Comments (C2,C3,C5,C6,C7) of Reviewer ehBt

Thanks for the comment.

C2:Quantitative validation of regime detection & EEG connectivity: Fig. 3 (main paper) links high qtq_t to frontal/temporal connectivity alterations (sec. 5.1). Tab 1 shows qtq_t rises from 0.572 pre-switch to 0.699 at transition and returns to 0.574 post-switch; $$\Deltaq=0.118[0.067,0.173] q=0.118[0.067,0.173], one-sided paired Wilcoxon p<0.001p<0.001, with AUROC 0.8920.892. Memory movement increases with qtq_t (0.392→0.441→0.503).

Tab 1:Quantitative validation of learned regime score (qtq_t) on DS4 dataset. Pre-switch:resting state, transition:first mental-arithmetic (MA) snapshot, and post-switch: the remaining MA snapshots. Values are summarized within subject and reported as median [IQR] across 36 paired subjects. (Δ\Delta q=qtransition−12(qpre+qpost)q=q_{\mathrm{transition}}-\frac{1}{2}(q_{\mathrm{pre}}+q_{\mathrm{post}}))

MetricLow qtq_t (<0.45)Medium qtq_t (0.45≤qtq_t<0.55)High qtq_t (≥0.55)Overall
A. Distribution of snapshots across qtq_t bins
All snapshots, n (%)115 (20.1%)149 (26.0%)308 (53.8%)572
Pre-switch, n (%)50 (19.9%)66 (26.3%)135 (53.8%)(q$_{pre}$=0.572) [0.524, 0.626]
Transition, n (%)2 (5.6%)6 (16.7%)28 (77.8%)qq_{transition}=0.699=0.699 [0.653, 0.752]
Post-switch, n (%)63 (22.1%)77 (27.0%)145 (50.9%)qq_{post}=0.574=0.574 [0.511, 0.621]
B. Bin-wise regime-score and memory behavior
qtq_t, median [IQR]0.368 [0.342, 0.389]0.514 [0.504, 0.525]0.660 [0.642, 0.807]0.576 [0.528, 0.616]
Memory movement ∥mt−mt−1∥2\|m_t-m_{t-1}\|_2, median [IQR]0.392 [0.311, 0.424]0.441 [0.431, 0.484]0.503 [0.472, 0.541]Overall: 0.471 [0.316, 0.542]; Transition/stationary: 2.102 [1.066, 2.454]
C. Evaluation against the known DS4 switch
Transition elevationΔq=0.118\Delta q=0.118 [0.067, 0.173]——p<0.001p<0.001
Transition vs. stationary discriminationAUROC = 0.892 [0.733, 0.987]———
Within-subject qtq_t variabilitySD = 0.127 [0.107, 0.170]Range = 0.439 [0.356, 0.531]—(qt∈[0.45,0.55](q_t\in[0.45,0.55]): 18.8 [12.5, 37.5]%

C3 :Stability and Contraction: A1 (in proof) holds for fixed parameters but not uniformly during training; utu_t is unnormalized. A2--A3 hold for Lipschitz MLPs. Boundedness prevents memory blow-up without κ<1\kappa<1, while contraction requires uncertified 2KLα<δ2KL_\alpha<\delta. Fig.4 shows stabilization of refinement terms and objective across GFP peaks, not direct contraction.

C5 Cross dataset/run EEG experiment: As DS3 has multiple runs within one session, we used strict leave-one-run-out evaluation. Tab 2 shows robust DS3 cross-run and DS7→DS8/DS9 zero-shot performance under temporal and cross-dataset shifts.

Tab 2: Cross-run/dataset robustness evaluation.

EvaluationTrain → TestAcc. (%)F1 (%)
Cross-run (DS3; held-out run)Runs 1,2,3 → Run 479.45 ± 3.3477.40 ± 2.46
Runs 1,2,4 → Run 383.33 ± 4.0182.76 ± 4.11
Runs 1,3,4 → Run 280.88 ± 4.0779.33 ± 4.23
Runs 2,3,4 → Run 184.11 ± 3.1283.88 ± 2.34
Cross-dataset (zero-shot; no target fine-tuning)DS7 → DS888.96 ± 2.3487.33 ± 3.02
DS7 → DS992.06 ± 2.7791.88 ± 3.45

C6: Fig.2 issue We need to clarify that Fig. 1 illustrates the proposed framework, and Fig. 2 presents the regime-aware memory analysis.

C7: Comparison with SOTA methods: Tab 3 compares RAGMO-Net with SALoM, EvoBrain, and FreeDyG. RAGMO-Net achieves the best Accuracy and F1 on EEG(Exp DS, DS3) and non-EEG datasets (DS7,DS8), confirming competitiveness against recent dynamic graph methods.

Tab 3: Comparison with SOTA methods

ModelExp DS (Acc)Exp DS (F1)DS3 (Acc)DS3 (F1)DS7 (Acc)DS7 (F1)DS8 (Acc)DS8 (F1)
SALoM [1]91.23 ± 2.3290.77 ± 1.7882.78 ± 3.7281.44 ± 3.2386.22 ± 2.3285.18 ± 3.7985.77 ± 4.1284.81 ± 3.31
EvoBrain [2]84.44 ± 3.2183.72 ± 2.7772.33 ± 4.1271.98 ± 3.6785.23 ± 4.0383.91 ± 3.4583.77 ± 2.4782.42 ± 3.42
FreeDyG [3]78.36 ± 3.1177.12 ± 4.2177.32 ± 5.2776.82 ± 6.0182.57 ± 3.7981.55 ± 4.3580.11 ± 2.3481.77 ± 1.89
RAGMO-Net97.64 ± 1.1296.73 ± 1.3488.48 ± 2.4587.56 ± 2.1192.83 ± 1.0190.18 ± 1.8793.29 ± 2.3492.55 ± 3.12

[1]:Hanwen Liu, et.al. Salom: Structure aware temporal graph networks with long-short memory updater. Advances in Neural Information Processing Systems, 2026.

[2]: Rikuto Kotoge, et.al. Evobrain: Dynamic multi-channel eeg graph modeling for time-evolving brain networks. Advances in Neural Information Processing Systems,2026

[3]:Yuxing Tian, et.al. Freedyg: Frequency enhanced continuous-time dynamic graph model for link prediction. In International Conference on Learning Representations, 2024

AuthorsOpenReview ↗

Other comments (C1, C4) of reviewer ehBt

C1:Methodological Novelty of the Scalar-Gated Affine Memory Update: We agree that Eqs.(5)--(6) resemble conventional gated memory updates. The novelty is not the gate itself, but its integration into a regime-aware, optimization-consistent dynamic-graph memory. RAGMO-Net constrains (Atm=ρtI)(A_t^m=\rho_t I) and (Btm=(1−ρt)Φ(ut)(B_t^m=(1-\rho_t)\Phi(u_t)), using a scalar retention factor rather than GRU/LSTM-style matrix-valued recurrence, reducing complexity as reflected in Table 4. Importantly, (ρt)(\rho_t) is coupled to regime score qtq_t, enabling explicit adaptation to stable/changing regimes. Unlike adaptive memory architectures that rely on learned prediction--correction or long--short memory modules, RAGMO-Net applies the closed-form regime-dependent proximal refinement in Eqs.(7)--(9), combining online-gradient, Bregman-proximity, and stability terms. Lemma 1 establishes boundedness margin, while Theorem 1 gives BIBO stability and conditional contraction. Table 3 shows 19.07, 9.20, and 12.29% accuracy drops without regime score, refinement, and stability penalty, showing gains arise from their combination rather than the scalar gate alone.

C4: Regime score & Biomarkers: Tab 1 (Biomarker 1) links qtq_t to frontotemporal connectivity biomarker (CFTC_{\mathrm{FT}}): median CFTC_{\mathrm{FT}} rises 0.072→0.127→0.129 across low/medium/high qtq_t. High-qtq_t states are more frequent in FTD vs HC (24.3% vs 8.0%) and distinguish FTD/HC (Mann--Whitney p=0.00235p=0.00235; AUROC=0.863, p=0.002p=0.002). High-vs-Low qtq_t also differs (paired Wilcoxon p=0.004p=0.004). These findings are consistent with EEG work showing frontotemporal connectivity provides discriminative information for FTD [1]. Tab.2 (Biomarker 2) links higher qtq_t to stronger network organization:node strength (0.499→0.634; p=0.039p=0.039), clustering coff(0.332→0.492; p=0.037p=0.037), and global efficiency (0.346→0.612; p=0.029p=0.029), with significant Holm-corrected Low-vs-High differences. These trends indicate stronger interaction, local organization, and global integration, supporting disorder-relevant network dynamics beyond one connectivity biomarker.

Tab 1: Relationship between learned regime score (qtq_t) and frontotemporal connectivity biomarker (CFTC_{FT})

MetricLow qtq_t (<0.45)Medium qtq_t (0.45≤qtq_t<0.55)High qtq_t (≥0.55)Overall
A. Distribution across qtq_t bins
All snapshots, n (%)161 (0.1%)117,889 (85.5%)19,797 (14.4%)137,847 snapshots
FTD snapshots, n (%)93 (0.2%)40,767 (75.6%)13,085 (24.3%)53,945 FTD snapshots
HC snapshots, n (%)68 (0.1%)77,122 (91.9%)6,712 (8.0%)83,902 HC snapshots
B. Frontotemporal connectivity biomarker
qtq_t, median [IQR]0.432 [0.431, 0.437]0.511 [0.508, 0.515]0.611 [0.600, 0.615]—
(CFT(C_{FT}), median [IQR]0.072 [0.051, 0.088]0.127 [0.117, 0.138]0.129 [0.115, 0.137]—
FTD (CFT(C_{FT}), median [IQR]0.087 [0.070, 0.092]0.136 [0.124, 0.139]0.134 [0.122, 0.138]High-(qt)(q_t) FTD vs. HC: p=0.00235p=0.00235
HC (CFT(C_{FT}), median [IQR]0.051 [0.036, 0.065]0.122 [0.115, 0.132]0.107 [0.075, 0.107]High-(qt)(q_t) AUROC = 0.863
C. Relationship with learned regime dynamics
High-qtq_t FTD vs. HCSubject-level frontotemporal connectivity——AUROC = 0.863; p=0.002p=0.002
High-qtq_t vs. Low-qtq_tPaired subject-level frontotemporal connectivity——p=0.004p=0.004

Tab 2: Relationship between learned regime score qtq_t and graph-network biomarkers

MetricLow qtq_t (<0.45)Medium qtq_t (0.45≤qtq_t<0.55)High qtq_t (≥0.55)Overall
A. Distribution across qtq_t bins
All valid snapshots, n (%)161 (0.1%)117,889 (85.5%)19,797 (14.4%)137,847 snapshots
FTD snapshots, n (%)93 (0.2%)40,767 (75.6%)13,085 (24.3%)53,945 FTD snapshots
HC snapshots, n (%)68 (0.1%)77,122 (91.9%)6,712 (8.0%)83,902 HC snapshots
B. Graph-network biomarkers
qtq_t, median [IQR]0.432 [0.431, 0.437]0.511 [0.508, 0.515]0.611 [0.600, 0.615]—
Mean node strength0.499 [0.479, 0.515]0.612 [0.573, 0.659]0.634 [0.589, 0.677]Friedman p=0.039p=0.039
Weighted clustering coefficient0.332 [0.279, 0.387]0.429 [0.384, 0.452]0.492 [0.429, 0.521]Friedman p=0.037p=0.037
Global efficiency0.346 [0.284, 0.391]0.501 [0.482, 0.545]0.612 [0.581, 0.642]Friedman p=0.029p=0.029
C. Low-qtq_t vs. High-qtq_t
Mean node strengthPaired subject-level comparison——$$p_{Holm}=0.025=0.025
Weighted clustering coefficientDO——$$p_{Holm}=0.037=0.037
Global efficiencyDO——$$p_{Holm}=0.041=0.041

[1]: Yuan Ma et.al. Classification of alzheimer’s disease and frontotemporal dementia using electroencephalography to quantify communication between electrode pairs. Diagnostics, 2024

Thank you for your reply. This has addressed my concerns. I will improve my score.

Reviewer_1K2YOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

RAGMO-Net combines graph representations, a learned regime score, scalar-gated memory, and a closed-form proximal refinement step for dynamic graph classification. The aim is to adapt memory retention to changing EEG connectivity. Experiments cover six EEG classification settings and five non-EEG benchmarks, with reported improvements over ten baselines, supported by ablations and runtime comparisons.

Critical strengths
  1. The memory update is simple and computationally attractive. The proximal subproblem has a valid closed-form solution, and the resulting update is a convex combination of previous memory and current representations. The bounded-state analysis and conditional contraction result help characterize the recurrence, although these guarantees concern a fixed network under the stated assumptions.

  2. The evaluation covers several EEG applications and non-EEG tasks, with comparisons against static graph, dynamic graph, and state-space models. Subject-disjoint EEG splits and validation-based early stopping are specified. Component and loss ablations provide evidence for the contribution of individual design choices, while runtime measurements help assess the practical benefit of the lightweight memory.

Critical weaknesses
  1. The justification for using Equation (13) as a training loss is incomplete. Lemma 1 establishes a lower bound for fixed network parameters, but that bound depends on learned feature scale. For illustration, setting m0=0m_0=0, ρ=q=0.5\rho=q=0.5, Φ(u)=0.2u\Phi(u)=0.2u, and η=λ=0.01\eta=\lambda=0.01 gives a first-step loss contribution of approximately −8.82×10−4∥u∥2-8.82\times10^{-4}\|u\|^2, which decreases as feature scale grows. It is unclear whether the stated weight decay or any feature normalization sufficiently controls this behavior during training. Fixed-network stability alone therefore does not justify the auxiliary loss.

  2. Important parts of the prediction problem are underspecified. Section 3.1 constructs ERP-peak graphs from averaged epochs, but the primary cohort contains resting-state EEG; epoch construction, averaging, and correlation windows are not explained. For DS6–DS10, the adaptation from graph-sequence classification to node, interaction, and user prediction is also unclear. These gaps affect what information enters each prediction and how the reported performance should be interpreted.

Questions for authors
  1. What controls the scale of learned graph states and excludes the behavior described above? Please specify the encoder normalization and the coefficients of the two auxiliary losses.

  2. How are the 17 ordered graphs constructed from resting-state recordings, and is accuracy measured per subject or per segment? Figure 4 shows 17 points for the healthy group and 19 for the dementia group despite the stated T=17T=17. What explains this discrepancy?

  3. How are prediction targets represented on DS6–DS10? Please clarify the chronological splits, class and negative sampling, and F1 averaging. For the TGB datasets, how does the evaluation relate to the standard candidate sets and MRR protocol?

6 comments
AuthorsOpenReview ↗

Comment 1 (C1) of Reviewer 1K2Y

C1: Theoretical Justification of Eq 13, Lemma 1, Training Stability and Auxiliary loss:

We thank the reviewer and agree. Lemma 1(c) (Eq 17) gives a lower bound for fixed network parameters only. The bound depends on K=max⁡{Ku,KΦ}K=\max\lbrace K_u, K_\Phi \rbrace, which varies with the learned encoder and injection-map parameters, so it gives no parameter-independent lower bound during training and does not by itself justify Eq 13 as an auxiliary loss.

First, we verified the example for η1>0\eta_1 > 0 (it vanishes for η1=0\eta_1 = 0). With m0=0m_0 = 0, ρ1,q1→0\rho_1, q_1 \to 0, Φ(u)=0.2u\Phi(u) = 0.2u and λ=0\lambda = 0, we get m~1=0.2u1\tilde m_1 = 0.2u_1 (Eqs 5–6), g1=m~1−u1=−0.8u1g_1 = \tilde m_1 - u_1 = -0.8u_1 (Eq 7), and m1=m~1−η1g1=(0.2+0.8η1)u1m_1 = \tilde m_1 - \eta_1 g_1 = (0.2 + 0.8\eta_1)u_1 (Eq 9). In Eq 13, the online-gradient and proximity terms give −(0.16η1+0.64η12)∥u1∥22-(0.16\eta_1 + 0.64\eta_1^2)\Vert u_1 \Vert_2^2 and 0.32η12∥u1∥220.32\eta_1^2\Vert u_1 \Vert_2^2, and the stability term vanishes, so the first-step contribution is

-\left(0.16\eta_1 + 0.32\eta_1^2\right)\Vert u_1 \Vert_2^2, \tag{R1}

which decreases without bound as the feature scale grows.

Second, completing the square identifies the source of this term. Expanding 12∥m−m~t∥22\tfrac12\Vert m-\tilde m_t \Vert_2^2 and combining the terms linear in mm gives

\eta_t\langle g_t, m\rangle + \tfrac12\Vert m-\tilde m_t \Vert_2^2 = \tfrac12\Vert m \Vert_2^2 - \langle m, \tilde m_t\rangle + \eta_t\langle m, g_t\rangle + \tfrac12\Vert \tilde m_t \Vert_2^2 = \tfrac12\Vert m \Vert_2^2 - \langle m,\ \tilde m_t - \eta_t g_t\rangle + \tfrac12\Vert \tilde m_t \Vert_2^2 . \tag{R2}

Adding and subtracting 12∥m~t−ηtgt∥22\tfrac12\Vert \tilde m_t - \eta_t g_t \Vert_2^2 completes the square in R2, and expanding it gives 12∥m~t∥22−12∥m~t−ηtgt∥22=ηt⟨gt,m~t⟩−ηt22∥gt∥22\tfrac12\Vert \tilde m_t \Vert_2^2 - \tfrac12\Vert \tilde m_t-\eta_t g_t \Vert_2^2 = \eta_t\langle g_t,\tilde m_t\rangle - \tfrac{\eta_t^2}{2}\Vert g_t \Vert_2^2. Hence

\eta_t \langle g_t, m\rangle + \tfrac{1}{2}\Vert m - \tilde m_t \Vert_2^2 = \underset{\text{(A)}}{\underbrace{\tfrac{1}{2}\big\Vert m - (\tilde m_t - \eta_t g_t)\big\Vert_2^2}} + \underset{\text{(B)}}{\underbrace{\eta_t \langle g_t, \tilde m_t\rangle - \tfrac{\eta_t^2}{2}\Vert g_t \Vert_2^2}} \tag{R3}

In the example, m1=m~1−η1g1m_1 = \tilde m_1 - \eta_1 g_1, so (A) =0= 0, and (B) =η1⟨−0.8u1,0.2u1⟩−η122(0.64)∥u1∥22= \eta_1\langle -0.8u_1, 0.2u_1\rangle - \tfrac{\eta_1^2}{2}(0.64)\Vert u_1 \Vert_2^2 equals R1. Thus the negative value comes entirely from (B).

Term (B) does not affect the memory update in Eq 9: m~t\tilde m_t and gtg_t are computed beforehand from mt−1m_{t-1} and utu_t (Eqs 5–7), so (B) is constant in the optimization variable mm of Eq 8 and shifts only the objective's value, not its minimizer. Setting the gradient of Eq 8 in mm to 0 gives

\big(m - (\tilde m_t - \eta_t g_t)\big) + 2\lambda(1-q_t)(m - m_{t-1}) = 0, \tag{R4}

whose solution is exactly Eq 9; (B) does not appear. However, (B) depends on the network parameters through m~t\tilde m_t and gtg_t, and Lomd\mathcal L_{\mathrm{omd}} enters Eq 11; even with the stop-gradient on gtg_t, ηt⟨gt,m~t⟩\eta_t\langle g_t,\tilde m_t\rangle passes gradients through m~t\tilde m_t. Preserving the minimizer thus does not preserve the training objective or its gradients.

Third, no explicit normalization is applied to utu_t, which concatenates the encoder output, structural summary and positional embedding. Training uses weight decay 10−410^{-4}, λomd=10−2\lambda_{\mathrm{omd}} = 10^{-2} and λq=10−1\lambda_q = 10^{-1}. Weight decay penalizes but does not cap large weights, so it gives no parameter-independent bound on utu_t, hence on KK. The 6.50-percentage-point drop without Lomd\mathcal L_{\mathrm{omd}} (Table 3) supports its empirical usefulness, not a uniform lower bound.

Finally, applying R3 at m=mtm = m_t to Eq 13 gives

\mathcal L_{\mathrm{omd}} = \sum_{t=1}^{T}\Big[\underset{\text{(A)}}{\underbrace{\tfrac12\big\Vert m_t-(\tilde m_t-\eta_t g_t)\big\Vert_2^2}} + \lambda(1-q_t)\Vert m_t-m_{t-1} \Vert_2^2 + \underset{\text{(B)}}{\underbrace{\eta_t\langle g_t,\tilde m_t\rangle - \tfrac{\eta_t^2}{2}\Vert g_t \Vert_2^2}}\Big]. \tag{R5}

Removing (B) gives the alternative

\mathcal L_{\mathrm{omd}} = \sum_{t=1}^{T} \Big[ \tfrac{1}{2}\big\Vert m_t - (\tilde m_t - \eta_t g_t)\big\Vert_2^2 + \lambda(1 - q_t)\Vert m_t - m_{t-1} \Vert_2^2 \Big]. \tag{R6}

As a sum of squared norms, it is nonnegative for all parameters; by R4, it keeps the update in Eq 9, but changes the training objective. With ct=2λ(1−qt)c_t = 2\lambda(1-q_t), R4 gives mt−(m~t−ηtgt)=−ct(mt−mt−1)m_t - (\tilde m_t - \eta_t g_t) = -c_t(m_t - m_{t-1}), so

\mathcal L_{\mathrm{omd}} = \sum_{t=1}^{T}\lambda(1-q_t)\big[1 + 2\lambda(1-q_t)\big]\Vert m_t - m_{t-1} \Vert_2^2 , \tag{R7}

a regime-weighted memory-movement penalty. The online-gradient term no longer appears explicitly but still shapes mtm_t through −ηtgt-\eta_t g_t in Eq 9. Here, Lomd\mathcal L_{\mathrm{omd}} is the auxilary loss.

Accordingly, we withdraw the claim that Lemma 1(c) justifies Eq 13 during end-to-end training and restrict it to its fixed-parameter scope.

AuthorsOpenReview ↗

Comments 2,3,4 (C2,C3, C4) of Reviewer 1K2Y

C2: EEG Preprocessing, Prediction-Target Construction, and Evaluation Protocol for DS6–DS10: For EEG datasets, recordings are bandpass-filtered to 0.5--40 Hz and segmented into 4-s epochs with 50% overlap. Epochs are averaged within each subject/channel at corresponding within-epoch time points. GFP is computed as the across-channel standard deviation of the averaged waveform, with peaks separated by at least 80 ms. For each peak, an 80-ms correlation window (±40\pm40 ms) at the same within-epoch position across retained epochs constructs a connectivity graph. The ordered peak-wise graphs form the subject-level dynamic sequence, so prediction uses temporal connectivity evolution rather than a single graph. Here, EEG Accuracy and F1-score are computed across held-out subjects, with each subject counted once. For DS6--DS10, the RAGMO-Net memory mechanism is unchanged; only graph-sequence construction and the prediction head are adapted to the task-specific prediction unit, using only historical information available before prediction time. For DS6, the task is node-level anomaly detection. A temporal graph sequence is constructed for each target node from historical interactions, graph-state/memory representations are aggregated, and one probability is produced per target node. Evaluation uses LOOCV over target nodes with chronological ordering to avoid leakage. For DS7--DS9, the prediction target is a future temporal interaction. Historical interactions are grouped into ordered temporal graph sequences while preserving source and candidate-destination roles. Each positive interaction is paired with one uniformly sampled destination-replaced negative, giving a 1:1 positive/negative ratio, and candidates are scored using the temporal graph-memory representation. For DS10, the prediction unit is the user. Each user is represented by a temporal sequence of graph states from historical interactions, and RAGMO-Net produces one binary user-level prediction; inverse-frequency class weighting is used during training. For DS7--DS9, Accuracy and F1 use this binary candidate-pair evaluation rather than the standard TGB candidate-set MRR protocol. MRR is not reported because the objective is binary interaction classification under the common non-EEG evaluation interface. F1 is computed over all predictions within each split and summarized as mean ±\pm standard deviation across splits. Overall, prediction units are subject-level for EEG, node-level for DS6, interaction-level for DS7--DS9, and user-level for DS10.

C3 Scale Control,Encoder Normalization and Auxiliary-Loss: No explicit normalization is applied to the encoder state utu_t. The classifier representation is zt=[ut;mt]z_t=[u_t;m_t], and training uses weight decay 10−410^{-4}; however, weight decay does not provide a parameter-independent bound on the learned feature scale. Lemma 1 guarantees bounded memory only for fixed network parameters under the assumption ∥ut∥2≤Ku\|u_t\|_2\leq K_u, and therefore does not by itself exclude feature-scale growth during training. We consequently restrict the boundedness claim to its fixed-parameter scope. In Eq. (11), the coefficient values of the two auxiliary losses are λ_omd = 10⁻² and λ_q = 10⁻¹. We additionally use η0=10−2\eta_0=10^{-2} (ηt=η0/t\eta_t=\eta_0/\sqrt{t}) for the proximal step-size schedule and λ=10−2\lambda=10^{-2} for temporal stability in Eqs.(8)--(9).

C4 Construction of ordered EEG Graphs, Subject-Level Accuracy, and Figure 4: The EEG graph construction is described in our response to Comment 2 (C2). Please refer to that response for details.

Accuracy measurement: As described in Section 4.2, evaluation is subject-wise, and all data from a participant remain within the same fold. The ordered graph sequence of each held-out subject produces one final subject-level class probability after temporal pooling and classification. Accuracy and F1-score are therefore computed across held-out subjects, with each participant contributing one prediction.

Figure 4 discrepancy: We agree that the current figure is inconsistent with the stated T=17T=17. Figure 4(a) shows 17 temporal positions, whereas Figure 4(b) includes two additional plotted positions because of plotting issue. We will correct Figure 4(b), its caption, and the corresponding indexing description to show exactly 17 positions for both groups.

Dear Reviewer 1K2Y,

Thank you for your time and thoughtful comments. As the discussion period is drawing to a close, we would be grateful if you could review our responses to your comments. We welcome any further questions or concerns and would appreciate the opportunity to address them before the discussion ends.

Best regards,

The Authors

Reviewer_vwyVOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework called "Regime-Aware Graph Memory Optimization Network (RAGMO-Net)," which adapts graph memory updates based on latent temporal regimes—distinct states or domains—that correspond to transitions in the system's condition. RAGMO-Net introduces regime-conditioned affine graph memory cells, effectively bridging dynamic graph learning with non-stationary optimization. Experiments conducted on both EEG and non-EEG datasets demonstrate that RAGMO-Net consistently outperforms existing, high-performing dynamic graph learning methods.

Critical strengths
  1. This method employs a proximal update—a variant of the online mirror descent algorithm—for updating the memory. This allows the analysis to leverage existing results concerning behavior under time-varying objective functions. Furthermore, it offers the advantage that each step can be computed in closed form. The memory transition is also linear, which facilitates analysis. Taking advantage of this analytical tractability, the stability and contractivity of the memory are established (Theorem 1), providing useful insights into the properties of the method.

  2. In this paper, the authors conduct detailed ablation experiments on a single dataset (inferred from the numerical values ​​to be the "Exp D" dataset). They provide a granular analysis of how the inclusion or exclusion of specific components, optimization strategies, and loss functions affects the results. By demonstrating the contribution of each part of the method—at least within the context of this specific dataset—they offer partial evidence suggesting the overall effectiveness of their approach.

Critical weaknesses
  1. Insufficient justification and evaluation of the memory update design: The primary contribution of this paper, as I understand it, lies in incorporating state changes—inherent in time-series data—into the memory update process. The update rule itself consists of a change gate and a linear transition; while this approach is tractable, it is also relatively simple and somewhat intuitive. The authors justify the use of a linear transition by noting that its behavior has been well-analyzed in existing literature. However, this merely speaks to the ease of analysis; it does not establish the validity of using such a transition for the specific problem at hand, nor does it demonstrate any advantage over alternative memory update mechanisms. In the context of memory-based methods—such as TGN—numerous approaches have been proposed that employ more sophisticated memory update schemes, many of which utilize non-linear updates (e.g., GRUs). Furthermore, no theoretical justification is provided to suggest that these non-linear updates are incapable of replicating the behavior of a linear transition equipped with a change gate. While the experimental evaluation includes comparisons with EEGMamba and DG-Mamba—both of which also rely on linear transitions for state updates—these comparisons are conducted at the level of the entire model, failing to isolate the specific contribution of the memory update rule itself. Moreover, among methods that maintain and sequentially update node-specific memory, the only comparison provided is against TGN; effectively, the only non-linear memory update mechanism considered is the GRU used within TGN. Unless the primary objective of this work is the prediction or classification of the state changes themselves, the focus remains on improving overall accuracy. Consequently, without comparing the proposed method against other approaches that also feature refined memory update mechanisms, it is difficult to accurately assess the effectiveness of the authors' specific memory update design.

  2. Concerns Regarding the Evaluation Protocol and Reproducibility: Regarding the evaluation protocol, several issues raise concerns about the reproducibility of the evaluation and the reliability of the reported performance: for the EEG datasets (DS1–DS5), while subject-based data splitting prevents information leakage, the small number of validation subjects (only a few) casts doubt on the stability of early stopping; furthermore, the procedures for hyperparameter selection and the aggregation unit for test metrics (i.e., whether calculated per graph or per subject) are not specified. For the non-EEG datasets (DS6–DS10), the description mentions both leave-one-out cross-validation (LOOCV) and time-series splitting—approaches that are inherently incompatible—making it impossible to determine the actual splitting method or the number of folds used; additionally, the lack of information regarding validation sets and training configurations makes it impossible to verify whether model selection was independent of the test period. Finally, given that the splitting methods differ between the two parts of the study, it is unclear whether the baselines were re-executed under identical conditions.

Questions for authors
  1. Could you provide a comparison with other memory-based methods? This could be a direct comparison, or one in which you replace the proposed method's memory update mechanism with those from existing approaches. Examples of baselines include PRES (ICLR 2024) and SALoM (NeurIPS 2025).

  2. Could you provide a more detailed explanation of how the theoretical results presented in Lemma 1 and Theorem 1 contribute to the effectiveness of the proposed method? This should include, for instance, an explanation of how these specific properties can lead to improvements in accuracy.

  3. Could you provide more clarity regarding the experimental setup? In particular, how were the hyperparameters determined?

4 comments
AuthorsOpenReview ↗

Comments 1 and 2 of Reviewer vwyV

Thanks for the comments.

C1:Justification,evaluation of the memory update design and comparison with recent memory-based methods: We agree that analytical tractability alone does not establish the suitability of the affine memory transition, nor do we claim that nonlinear recurrent updates such as GRUs cannot represent similar dynamics. RNNs and GRU-based architectures have strong approximation capabilities [1,2]. Our motivation is instead to introduce an explicit regime-conditioned inductive bias for non-stationary dynamic graphs. In Sec. 3.1, a learned scalar retention factor conditioned on the graph representation and regime score explicitly controls retain--adapt behavior, while Sec. 3.2 further refines the provisional memory using the non-stationary objective. Lemma 1 and Theorem 1 establish boundedness and contraction, rather than superiority over nonlinear recurrent mechanisms. As suggested, we compare RAGMO-Net with PRES [3] and SALoM [4] in Table 1. For a fair comparison, PRES and SALoM were evaluated using the same dataset-specific protocol as RAGMO-Net: subject-wise stratified 5-fold CV for EEG datasets (Exp DS, DS3) and LOOCV over task-specific prediction units for non-EEG datasets (DS7, DS8). The same train/validation/test partitions and evaluation metrics were used across models, with validation data used only for model selection. RAGMO-Net achieves higher Accuracy/F1 across EEG and non-EEG datasets. Table 3 of the main paper further shows that removing the regime score degrades performance, while removing proximal refinement reduces accuracy from 97.64% to 88.44%. These results support the effectiveness of regime-conditioned memory adaptation and optimization-based refinement beyond the overall architecture.

Table 1: Comparison with recent memory-based methods. Results(%) = mean ± standard deviation. EEG datasets-Exp DS and DS3, Non-EEG datasets- DS7 and DS8

ModelExp DS (Acc)Exp DS (F1)DS3 (Acc)DS3 (F1)DS7 (Acc)DS7 (F1)DS8 (Acc)DS8 (F1)
SALoM[4]91.23 ± 2.3290.77 ± 1.7885.78 ± 3.7284.44 ± 3.2387.22 ± 2.3286.18 ± 3.7983.77 ± 4.1282.81 ± 3.31
PRES[3]86.23 ± 2.3785.77 ± 3.2381.22 ± 2.3381.77 ± 1.9782.45 ± 3.7582.29 ± 4.0287.78 ± 4.2387.90 ± 3.72
RAGMO-Net97.64 ± 1.1296.73 ± 1.3488.48 ± 2.4587.56 ± 2.1192.83 ± 1.0190.18 ± 1.8793.29 ± 2.3492.55 ± 3.12

C2: Experimental setup, Hyperparameter selection, Evaluation Protocol and Reproducibility: For EEG datasets (Exp DS and DS1--DS5), we use subject-wise stratified 5-fold CV. In each fold, one subject split is held out for testing, while 30% of the remaining subjects per class are used for validation, with at least two validation subjects per class. All graphs from each subject remain in the same fold to prevent leakage. To reduce sensitivity to small validation sets, we use the same training configuration and fixed early-stopping settings across folds: BCE loss, AdamW with learning rate 3×10−43\times10^{-4} and weight decay 10−410^{-4}, batch size 16, up to 150 epochs, patience 12, and minimum improvement 10−410^{-4}. Results are reported as mean$\pm$SD over five folds. Hyperparameters are selected using validation performance only. Specifically, η0=10−2\eta_0=10^{-2} is selected on validation data with ηt=η0/t\eta_t=\eta_0/\sqrt{t}; sensitivity to temporal regularization λ\lambda and attention dimension is reported in Appendix D.1, with λ=10−2\lambda=10^{-2} and dimension 64 performing best. For non-EEG datasets (DS6--DS10), we use LOOCV over task-specific prediction units rather than per-instance LOOCV: target nodes for DS6, candidate temporal interactions for DS7--DS9, and users for DS10. Within each evaluation split, strict chronology ensures training uses past interactions and evaluation uses future interactions, preventing temporal leakage. Thus, LOOCV defines the held-out prediction unit, while chronology defines the temporal information boundary; they are complementary rather than conflicting. Results are reported as mean$\pm$SD over evaluation splits. All baselines were re-executed using the same dataset-specific partitions, preprocessing, held-out units, and evaluation metrics as RAGMO-Net, with model selection performed without test-set access.

[1]:Chang hoon Song et.al., Minimal width for universal property of deep rnn. Journal of Machine Learning Research, 24(121):1–41, 2023.

[2]:Aleksandar Petrov et.al. Universal in-context approximation by prompting fully recurrent models. Advances in Neural Information Processing Systems, 37:72061–72093, 2024.

[3]: Junwei Su, et.al. Pres: Toward scalable memory-based dynamic graph neural networks. arXiv preprint arXiv:2402.04284, 2024

[4]: Hanwen Liu, et.al. Salom: Structure aware temporal graph networks with long-short memory updater. Advances in Neural Information Processing Systems, 38:22843–22871, 2026.

AuthorsOpenReview ↗

Comment 3 of Reviewer vwyV

C3: Theoretical Contribution of Lemma 1 and Theorem 1 to Model Effectiveness and Accuracy:

Thanks for the insightful comment. Lemma 1 and Theorem 1 explain how bounded and stable memory dynamics can support reliable classification. They provide a mechanism that can preserve predictions under suitable perturbations, while the accuracy gains themselves are demonstrated empirically.

First, Lemma 1 ensures bounded representations. Since ∥ut∥2,∥mt∥2≤K\Vert u_t \Vert_2, \Vert m_t \Vert_2 \le K, the joint state satisfies ∥zt∥2≤2K\Vert z_t \Vert_2 \le \sqrt{2}K. Attention pooling, being a convex combination, preserves this bound. Thus, recurrent accumulation cannot make the representation supplied to the classifier grow without bound. With fixed network parameters and Lipschitz attention-scoring and classification modules, the complete attention-and-classification mapping is Lipschitz on this bounded state domain.

Second, Theorem 1 connects stability to preservation of classification decisions. For two input sequences satisfying the assumptions of Theorem 1 with identical initialization, under the contraction condition κ<1\kappa < 1,

max⁡t≤T∥mt−mt′∥2≤εmax⁡1−κ,\max_{t \le T} \Vert m_t - m_t' \Vert_2 \le \frac{\varepsilon_{\max}}{1 - \kappa},

where εmax⁡=max⁡k≤Tεk\varepsilon_{\max} = \max_{k \le T} \varepsilon_k is the maximum one-step input sensitivity defined in Theorem 1, accounting for changes in both graph states and structural-drift descriptors. Let Du=max⁡t≤T∥ut−ut′∥2D_u = \max_{t \le T} \Vert u_t - u_t' \Vert_2. For fixed positional indices, the logit difference satisfies

∣f−f′∣≤LreadDu2+(εmax⁡1−κ)2=:B,|f - f'| \le L_{\mathrm{read}} \sqrt{D_u^2 + \left(\frac{\varepsilon_{\max}}{1 - \kappa}\right)^2} =: B,

where LreadL_{\mathrm{read}} is the Lipschitz constant of the attention-and-classification mapping on the bounded state domain, using the maximum per-step Euclidean norm. A correctly classified subject with label y∈{0,1}y \in \lbrace 0, 1 \rbrace therefore remains correctly classified under a label-preserving perturbation whenever (2y−1)f>B(2y - 1) f > B. This includes artifact-induced changes in the encoded graph sequence with peak indices held fixed, provided the assumptions hold. This applies the Lipschitz-margin principle used in robustness certification [1].

Third, regime-aware refinement balances retention and adaptation. Equation (9) gives

mt−mt−1=m~t−ηtgt−mt−11+2λ(1−qt).m_t - m_{t-1} = \frac{\tilde m_t - \eta_t g_t - m_{t-1}}{1 + 2\lambda(1 - q_t)}.

For a fixed numerator, low regime scores attenuate memory movement, helping preserve history during stable periods, while high scores relax this constraint and permit adaptation. Because the gates also depend on the input and memory, this motivates a stability–adaptation trade-off rather than proving that regime awareness always reduces overall sensitivity.

Fourth, the experiments support these components' effectiveness. Table 3 (in the main paper) shows accuracy drops of 12.29, 19.07 and 9.20 percentage points when removing the stability penalty, the regime score and the proximal refinement, respectively. Figure 5(a) shows the best measured accuracy among the tested settings at λ=10−2\lambda = 10^{-2}, consistent with balancing retention and adaptation.

Finally, these observations are consistent with related work. Orvieto et al. [2] demonstrate the effectiveness of carefully parameterized linear recurrent units for long-sequence modeling. Zucchet and Orvieto [3] analyze parameter sensitivity in long-memory recurrent networks, while Su et al. [4] connect memory coherence with training convergence in memory-based dynamic GNNs. These studies provide complementary motivation for controlling memory dynamics. We will clarify our theoretical connection to classification robustness in the revised paper.

[1] Alexandre Araujo et al. A unified algebraic perspective on Lipschitz neural networks. In International Conference on Learning Representations, 2023.

[2] Antonio Orvieto et al. Resurrecting recurrent neural networks for long sequences. In International Conference on Machine Learning, PMLR, 2023.

[3] Nicolas Zucchet and Antonio Orvieto. Recurrent neural networks: Vanishing and exploding gradients are not the end of the story. In Advances in Neural Information Processing Systems, 2024.

[4] Junwei Su et al. PRES: Toward scalable memory-based dynamic graph neural networks. In International Conference on Learning Representations, 2024.

Dear Reviewer vwyV,

Thank you for your time and thoughtful comments. As the discussion period is drawing to a close, we would be grateful if you could review our responses to your comments. We welcome any further questions or concerns and would appreciate the opportunity to address them before the discussion ends.

Best regards,

The Authors

Dear Reviewer,

As the rebuttal period is coming to an end, I would like to kindly ask you to read the authors' response and reply as soon as possible.

Thank you for your prompt attention.

Best regards,

Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.