acceptodds
Under review as a conference paper at ICLR 2027

Near-Balanced Signed Graph Learning & Filtering for Depression Score Prediction

Abstract

Recent studies have shown that strongly anti-correlated signals often induce nearly balanced signed graphs containing only a small number of inconsistent negative cycles (frustrations). However, existing methods do not learn such statistics as near-balanced signed graphs nor exploit this graph structure for signal filtering. We formulate an optimization that jointly learns a near-balanced signed graph and filters signals by alternating between balanced and frustrated graph components. Specifically, we construct a generalized expectation maximization (GEM) framework that alternates between graph learning (M-step) and signal estimation (E-step). The M-step learns a near-balanced graph by trading off data fidelity and graph frustration, while the E-step denoises a signal by minimizing a graph Laplacian regularizer, alternating between low-pass filtering on the balanced component and solving a linear system for the frustrated component. Unrolling the resulting iterations yields a lightweight and interpretable transformer-like neural network with direct correspondence to the underlying optimization steps. Applied to fMRI-based depression score prediction, the proposed framework matches state-of-the-art (SOTA) performance while reducing parameters to fewer than 5% of SOTA methods, outperforming both positive and perfectly balanced graph models.

Primary area:
optimization
Keywords:
Near-Balanced Signed Graph, Spectral Denoising, Generalized Expectation Maximization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_dUeKOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper learns a near-balanced signed graph and filters its balanced and frustrated components separately. It unrolls these steps into a network for depression score prediction and provides a spectral condition for near-balance.

Critical strengths
  1. Allowing some frustration is a reasonable extension of strictly balanced graphs, which restrict the relationships the graph can represent.

  2. The network is compact, with clear graph-learning and filtering components.

Critical weaknesses
  1. It is not clear why the proposed updates qualify as GEM. Equations (6)–(7) and (14)–(15) use a point estimate of the latent signal rather than the posterior expectation in the EM objective. The paper also does not show that improving the substituted objectives improves the original objective. This leaves the claimed connection to GEM, and its guarantees, unsubstantiated.

  2. The main methodological contribution needs a clearer justification. Splitting the Laplacian into balanced and frustrated components leaves the exact denoising solution unchanged. Any advantage must therefore come from how the split is implemented, such as finite unrolling or separate parameterization. I am unsure what this construction adds over directly filtering the full signed Laplacian.

Questions for authors
  1. What objective do the implemented updates improve, and why should they be viewed as GEM rather than alternating MAP optimization?

  2. Does Proposition 1 apply to the learned graph in Equation (8), or is it only motivation for the method?

  3. What fraction of the learned edge weight is frustrated, and how does the fully balanced baseline handle those edges?

2 comments
AuthorsOpenReview ↗

Response (1) to Reviewer dUeK

Response to Weakness 1:

We thank the reviewer for this comment. Our GEM formulation is based on the single negative log-posterior objective in (3), consisting of the data-fidelity term ∥y−x∥22\|\mathbf{y} - \mathbf{x}\|^2_2​, graph-smoothness term x⊤Lθx\mathbf{x}^\top \mathcal{L}_{\boldsymbol{\theta}} \mathbf{x}, and graph-prior term −log⁡det⁡†(Lθ)-\log \det^\dagger (\mathcal{L}_{\boldsymbol{\theta}}). Equations (14)-(15) update the latent signal x\mathbf{x} by solving the corresponding MAP problem exactly while fixing θ\boldsymbol{\theta}, thereby guaranteeing a non-increase in the objective (3). Equations (6)-(7) then update θ\boldsymbol{\theta} while fixing x\mathbf{x} through our near-balanced graph-learning procedure, which is likewise derived from the same objective.

We note that our E-step employs a point-estimate (MAP) update of the latent signal rather than an explicit posterior expectation. Our use of GEM therefore follows the generalized-EM viewpoint, where latent-variable updates need not take the form of exact posterior expectations. We agree that the connection between the update rules and the underlying objective (3), as well as the monotonicity properties of the resulting alternating optimization scheme, was not discussed explicitly in the submission. We will clarify this relationship and provide additional theoretical discussion in the revised manuscript.

Response to Weakness 2:

We thank the reviewer for this comment. We agree that, for the underlying optimization problem, decomposing the signed Laplacian into balanced and frustrated components does not change the exact denoising solution at convergence. However, the proposed method does not solve the optimization problem via an arbitrarily large number of iterations until convergence. Instead, we unroll a finite number of E- and M-steps into a transformer-like neural network with learnable parameters.

In this finite-depth setting, the parameterization of the filtering operation becomes important. The balanced/frustrated decomposition introduces additional structure that allows the resulting graph filters to adapt differently to balanced and frustrated interactions present in the data. Consequently, even though the asymptotic solution remains unchanged, the finite-layer network can achieve improved denoising and prediction performance through a more expressive and data-adaptive parameterization.

This distinction is discussed in Appendices F and G. In particular, the benefit of the proposed decomposition should be understood in the context of finite unrolling and learned graph filtering, rather than in terms of changing the ultimate optimization solution. Empirically, Table 2 supports this interpretation: the near-balanced model consistently outperforms both the positive-graph and fully-balanced alternatives under the same finite-depth architecture, demonstrating that explicitly modeling frustration provides a practical advantage beyond directly filtering with the full signed Laplacian.

Moreover, to directly evaluate the necessity of the balanced/frustrated decomposition, we conducted an ablation study using the same learned near-balanced graph Laplacian Lθ\mathcal{L}_\theta​ and the same three-layer unrolled architecture. If the decomposition were unnecessary, then directly filtering with the full signed Laplacian should perform comparably to the proposed split-filtering strategy under the same finite-depth computation budget. We compared three configurations:

  • Joint: filtering using the full near-balanced Laplacian Lθ\mathcal{L}_\theta​.
  • Split (CG/CG): filtering LθB\mathcal{L}_\theta^B​ and Fθ\mathbf{F}_\theta​ separately using CG for both components.
  • Split (Lanczos/CG): the proposed approach, using Lanczos approximation for LθB\mathcal{L}_\theta^B​ and CG for Fθ\mathbf{F}_\theta​. | Filter configuration | Reconstruction MSE ↓ | |:---|---:| | Joint | 0.14387786 | | Split (CG/CG) | 0.14395270 | | Split (Lanczos/CG) | 0.14215672 |

The results provide two insights. First, Joint and Split(CG/CG) perform similarly, confirming the reviewer's observation that decomposition alone does not substantially alter the underlying optimization problem. Second, Split(Lanczos/CG) achieves the lowest reconstruction MSE. This demonstrates that the value of the decomposition lies in enabling specialized processing of the balanced component. In particular, the balanced graph Laplacian LθB\mathcal{L}_\theta^B​ admits a meaningful low-frequency eigenspace associated with manifold harmonics, allowing a finite-step Lanczos approximation to efficiently capture the dominant signal subspace. Such a strategy is less effective when filtering directly with the full signed Laplacian. Therefore, the benefit arises not from changing the asymptotic solution, but from providing a more suitable parameterization for a finite-depth unrolled architecture.

AuthorsOpenReview ↗

Response (2) to Reviewer dUeK

Response to Question 1.

We thank the reviewer for this question. The updates are derived from the single negative log-posterior objective in (3), \begin{equation} | \mathbf{y} - \mathbf{x}|_2^2 + \mathbf{x}^\top \mathcal{L}_{\boldsymbol\theta} \mathbf{x} - \log\det^\dagger(\mathcal{L}_{\boldsymbol\theta}) \end{equation} where x\mathbf{x} is treated as a latent signal and θ\boldsymbol{\theta} parameterizes the graph model. The E-step updates x\mathbf{x} by solving the corresponding MAP estimation problem while fixing θ\boldsymbol{\theta}, and the M-step updates θ\boldsymbol{\theta} while fixing x\mathbf{x} through the proposed near-balanced graph-learning procedure. Thus, both update blocks are derived from and seek to improve the same underlying objective.

We agree that the resulting procedure can also be viewed as an alternating MAP optimisation scheme. Our use of the GEM terminology is motivated by the latent-variable interpretation of x\mathbf{x} and the alternating update of latent variables and model parameters under a common posterior objective. Moreover, the E-step employs a point-estimate (MAP) update of the latent signal rather than an explicit posterior expectation, placing the method closer to the generalized-EM viewpoint than to classical EM. We will revise the manuscript to clarify this connection and make explicit that the proposed algorithm may be viewed equivalently as a latent-variable alternating optimisation procedure arising from the negative log-posterior objective (3).

Response to Question 2.

We thank the reviewer for this question. Proposition 1 should primarily be viewed as a theoretical motivation for the near-balanced graph prior adopted in our M-step, rather than as a theorem directly characterizing the learned graph in Eq. (8). Specifically, Proposition 1 shows that under a natural statistical construction, thresholded affinity graphs tend to exhibit a small proportion of frustrated edges, leading to near-balanced rather than fully balanced or highly frustrated structures. This observation motivates our assumption that near-balanced signed graphs provide a suitable model class for capturing natural interaction patterns. The graph in Eq. (8), however, is not constructed via the thresholding process considered in Proposition 1. Instead, it is learned adaptively from node features through the graph-learning procedure in the M-step. Thus, Proposition 1 does not directly predict the structure of the learned graph.

Rather, the relationship is that Proposition 1 motivates the near-balanced prior, while Eq. (8) implements a task-adaptive graph-learning mechanism constrained by that prior. The effectiveness of this modelling choice is then evaluated empirically through the ablation studies in Table 2 and the additional experiments reported in our rebuttal. We will revise the manuscript to make the intended role of Proposition 1 more explicit.

Response to Question 3.

We thank the reviewer for this question. We collected statistics on the learned near-balanced graphs across the three unrolled neural layers. The balanced component accounts for approximately 82.78%–83.82% of all edges and 90.40%–91.45% of the total absolute edge-weight mass, implying that the frustrated component accounts for only 16.18%–17.22% of the edges and 8.55%–9.60% of the total absolute edge-weight mass.

These statistics are consistent with the near-balanced graph assumption underlying the proposed method: most interactions are captured by the balanced component, while a relatively small amount of frustration is retained. In the fully balanced baseline, frustrated edges are removed by enforcing exact structural balance. Consequently, the resulting graph preserves the dominant balanced structure but discards the frustration component. This explains why the fully balanced model achieves denoising MSE comparable to that of the near-balanced model, since the balanced component contains more than 90% of the edge-weight mass. However, as shown in our additional ablation study, the frustrated component contributes disproportionately to downstream prediction performance. Thus, enforcing full balance removes a comparatively small portion of the graph structure that nevertheless contains information important for BDI prediction.

We will include these statistics and their interpretation in the revised manuscript to better clarify the distinction between the near-balanced and fully balanced graph models.

Reviewer_8qT4OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes to use near balanced graphs to decompose signals into fustrated and balanced components and then use both components to filter signals. The paper then proposes to learn the balanced laplacian and the fustrated component as signals in a parameterized way via GEM through a pretext denoising task and then use the frozen learned parameters for a downstream task of predicting subject-level BDI scores from resting-state fMRI.

Critical strengths
  • The paper is clear
  • The experimental results are strong (though I have a question on other baselines' implementations)
  • Idea seems to build well on prior literature (though I have questions on the theoretical motivations)
Critical weaknesses
  • The paper does pretaining, then pca, then uses the pca to train on downstream task. What are the benefits of this pca step? do any other baseline forllow this same pipeline of pretraining? As it has been demonstrated already that pretraining benefits immensely this may be unfair comparison if no other baselines do
  • The theory looks okay, but can the authors explain or may be experimentally jsutify splitting the balanced composnent from the fustrated component? i.e., what is the benefit of splitting L in equation 16 into LB and F (which is the whole contribution of the paper) we will like to know if any prior baseline use the same equation 16 and unroling framwork but only used L directly? if not we will like to see those results as well (as that will enhance the contribution/motivation of the paper).
Questions for authors
  • Please see section on Critical weaknesses
2 comments
AuthorsOpenReview ↗

Response (1) to Reviewer 8qT4

Response to Weakness 1:

We thank the reviewer for raising this point. The purpose of the PCA step is to reduce the dimensionality of the learned window descriptors before fitting the downstream BDI predictor, since the number of independent training subjects in each LOSO fold is limited. In the audited implementation, the signal and graph descriptors are separately reduced to 8 and 12 principal components prior to regression.

Importantly, the denoiser is pretrained using only the self-supervised signal-reconstruction objective and does not use BDI labels. PCA and the downstream predictor are then fitted exclusively on the training subjects within each LOSO fold, preventing information leakage. Regarding fairness, the key issue is separating the effect of (i) self-supervised pretraining, (ii) PCA-based dimensionality reduction, and (iii) the proposed near-balanced graph model. We agree that additional experiments in which eligible baselines are given access to the same pretrained representations and downstream readout protocol would more cleanly isolate these factors. We will clarify the comparison protocol in the revised manuscript and, where feasible, include such controlled comparisons.

In the baselines we list in main results table, graphTRIP-BDI jointly trains a variational graph autoencoder (VGAE) and a prediction head using both graph reconstruction and score prediction objectives. Tolle et al. also report a separate transfer experiment in which they train the model on a source cohort, freeze the VGAE, and fine-tune only the prediction head on the target cohort. Tangent Ridge also fits its representation before training the predictor. It estimates a tangent-space reference from the training subjects without using outcome labels and then fits a ridge regressor on the resulting connectivity features.

AuthorsOpenReview ↗

Response (2) to Reviewer 8qT4

Response to Weakness 2:

We thank the reviewer for this comment on the rationale for splitting the signal filtering into balanced and frustrated components. This question is largely addressed in Appendix G, but we summarize the key intuition and provide an additional ablation study here. Given our assumption that the target signal resides on a near-balanced signed graph (motivated by Proposition 1), a significant portion of the signal is balanced and admits a one-to-one mapping to a signal on a corresponding positive graph (Section 2.3). A positive graph benefits from the conventional notion of graph frequencies and manifold harmonics, where a smooth signal can be well approximated using the low-frequency eigenspace of the balanced Laplacian LθB\mathcal{L}_\theta^B. Consequently, a truncated unrolled Lanczos implementation (Appendix H), with learned parameters optimized to approximate the low-frequency subspace of LθB\mathcal{L}_\theta^B​, can efficiently recover the balanced component of the signal.

The key point is that our method does not solve the signed-graph filtering problem to convergence. Instead, we unroll a finite number of neural layers with learnable parameters. In the limit of exact convergence, directly filtering with the full signed Laplacian Lθ\mathcal{L}_\theta​ or splitting it into LθB\mathcal{L}_\theta^B and Fθ\mathbf{F}_\theta​ would lead to the same denoising solution. However, under finite unrolling, the parameterization becomes important. The balanced/frustrated decomposition introduces additional structure that allows the architecture to process the dominant balanced component and the weaker frustrated component separately, enabling the learned parameters to better adapt to their distinct characteristics. Thus, the benefit arises not from changing the asymptotic solution, but from providing a more expressive and data-adaptive finite-depth implementation.

To directly evaluate the necessity of the proposed decomposition, we conducted an ablation study using the same learned near-balanced graph Laplacian Lθ\mathcal{L}_\theta​ and the same three-layer unrolled architecture. If the decomposition were unnecessary, then directly filtering the full signed Laplacian should perform comparably to the proposed split-filtering strategy under the same finite-depth computation budget. We compared three configurations:

  • Joint: filtering using the full near-balanced Laplacian Lθ\mathcal{L}_\theta​.
  • Split (CG/CG): filtering LθB\mathcal{L}_\theta^B and Fθ\mathbf{F}_\theta​ separately using conjugate gradient (CG) for both components.
  • Split (Lanczos/CG): the proposed approach, using Lanczos approximation for LθB\mathcal{L}_\theta^B​ and CG for Fθ\mathbf{F}_\theta​. | Filter configuration | Reconstruction MSE ↓ | |:---|---:| | Joint | 0.14387786 | | Split (CG/CG) | 0.14395270 | | Split (Lanczos/CG) | 0.14215672 |

The results provide two important insights. First, Joint and Split (CG/CG) achieve very similar performance, confirming the reviewer's observation that the decomposition alone does not fundamentally alter the underlying optimization problem. Second, the proposed Split (Lanczos/CG) strategy achieves the lowest reconstruction MSE. This demonstrates that the value of the decomposition lies in enabling specialized processing of the balanced and frustrated components.

In particular, the balanced component LθB\mathcal{L}_\theta^B​ corresponds to a positive graph whose low-frequency eigenspace captures the dominant smooth signal content. A finite-step Lanczos approximation can therefore efficiently target this subspace, whereas directly filtering the full signed Laplacian cannot exploit this structure as effectively. Hence, the benefit of the proposed decomposition is not that it changes the ultimate optimization solution, but that it enables a more suitable parameterization and approximation strategy for a finite-depth unrolled network.

Finally, recovery of the balanced component is important for denoising performance because it captures the dominant smooth signal structure. The proposed ADMM framework additionally recovers the frustrated component, which we observed to be important for downstream prediction. As shown in Table 2, the fully balanced graph achieves denoising MSE comparable to the near-balanced model, yet yields substantially inferior PCC. This suggests that the frustrated component contains discriminative information that is largely irrelevant for denoising but beneficial for depression-score prediction. The balanced/frustrated decomposition therefore enables the network to preserve and exploit this information rather than suppressing it through a purely balanced processing model.

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

This paper introduces a near-balanced signed graph learning and filtering framework for fMRI-based depression score prediction, where the graph Laplacian is decomposed into balanced and frustrated components and the resulting GEM iterations are unrolled into a lightweight network. The approach offers a principled way to retain controlled frustration in signed graph learning, together with an interpretable unrolling design that assigns explicit filtering semantics to the Lanczos parameters. Overall, the modeling perspective is interesting, but the central claims about near-balance and the necessity of the balanced/frustrated decomposition are supported only indirectly, and the lightweight and interpretability claims are not adequately substantiated.

Critical strengths
  1. Modeling anti-correlated signals through near-balanced signed graphs and decomposing the Laplacian into balanced and frustrated components offers a principled alternative to ignoring edge signs or enforcing full balance.
  2. Unrolling GEM iterations with explicit filtering semantics in the Lanczos parameters gives the architecture more interpretability than a black-box graph network.
Critical weaknesses
  1. The evaluation is limited to basic baseline comparisons and a simple graph-structure ablation, so the advantage of near-balanced graphs and the necessity of the balanced/frustrated decomposition are supported only indirectly through prediction metrics, without direct evidence on frustrated-edge proportions, polarity stability, or brain-interaction correspondence.
  2. Although the method assigns explicit semantics to learned polarities, frustrated edges, and filter responses, these quantities are never connected to brain structure or examined for whether they behave as the optimization semantics predict, leaving the interpretability claim without empirical support.
  3. The full-balanced graph reaches comparable denoising MSE but a sharply negative BDI PCC of -0.379, which is explained only briefly, and Proposition 1 concerns thresholded affinity graphs while the actual M-step builds graphs from learned features, with the relationship between the two left unquantified.
Questions for authors
  1. Can the authors provide direct evidence that near-balanced graphs capture anti-correlated structure better than full-balanced or positive graphs?
  2. Can the necessity of the balanced/frustrated decomposition be shown through component-wise ablations or alternative filtering strategies?
  3. What is the full pipeline parameter count including the downstream predictor under a matched comparison protocol?
  4. What explains the negative PCC of the full-balanced graph, and how does Proposition 1 relate quantitatively to the learned graph?
3 comments
AuthorsOpenReview ↗

Response (1) to Reviewer WMod

Response to Weakness 1 & 2:

We thank the reviewer for the comments regarding interpretability and validation of the learned graph structure. We believe both concerns arise from interpreting the learned near-balanced signed graphs as an estimate of true brain connectivity. This is not the claim made in our paper. Rather, the learned graphs are latent computational structures used within the proposed graph-learning and graph-filtering framework for signal denoising and downstream depression-score prediction.

Importantly, the proposed architecture contains multiple unrolled E- and M-steps and therefore learns multiple graph structures across layers. These graphs are not intended to recover a single anatomical or functional brain network. Instead, they serve as intermediate computational representations that parameterize the graph-filtering operations of the unrolled network. In fact, Appendix C shows that the learned graph-filtering layers in the E-step can be interpreted as a self-attention mechanism, thereby establishing a direct connection between the proposed unrolled architecture and transformer-style networks. From this perspective, the learned near-balanced signed graphs should be viewed as attention structures within an interpretable unrolled transformer, rather than estimates of physical brain connectivity. Consequently, there is no corresponding biological ground truth against which quantities such as polarities, frustrated edges, or graph balance can be validated, nor do we claim that these quantities correspond to anatomical or functional brain interactions.

The term “interpretable” is used only once in the manuscript (Introduction, paragraph 3), where it follows the standard definition in the algorithm-unrolling literature [1,2,3]. Specifically, an unrolled network is considered interpretable because each neural layer corresponds directly to one iteration of a well-defined optimization algorithm minimizing an explicit objective. This notion differs fundamentally from neuroscientific interpretability, where learned variables are expected to correspond to physical brain structures or mechanisms. Our interpretability claim concerns the former, not the latter.

Thus, the semantics associated with learned polarities, frustrated edges, and graph-filtering responses should be interpreted through the optimization and self-attention mechanisms implemented by the unrolled transformer, rather than through correspondence to brain anatomy or functional connectivity. The learned layers are not treated as black-box operations; they admit both an optimization-based interpretation through the GEM framework and an architectural interpretation through their relationship to self-attention. Regarding the necessity of the balanced/frustrated decomposition, this is evaluated through the graph-structure ablation in Table 2, where positive-graph, fully-balanced, and near-balanced graph models are compared under an identical experimental protocol. The near-balanced model achieves the best downstream prediction performance, while the fully-balanced model performs substantially worse despite comparable denoising accuracy. These results support the importance of explicitly modelling a small amount of frustration rather than enforcing strict balance.

We agree that additional analyses of internal quantities such as frustration levels or polarity stability may provide further insight into the learned model dynamics. However, such analyses should be viewed as diagnostics of the learned computational representation and self-attention structure within the unrolled transformer, rather than validation against biological brain connectivity. The latter is outside the scope of the claims made in this work.

References

[1] V. Monga, Y. Li, and Y. C. Eldar, “Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing,” IEEE Signal Processing Magazine, vol. 38, no. 2, pp. 18–44, 2021.

[2] Y. Yu, S. Buchanan, D. Pai, T. Chu, Z. Wu, S. Tong, B. Haeffele, and Y. Ma, “White-box transformers via sparse rate reduction,” NeurIPS, 2023.

[3] T. Thuc, P. Eftekhar, S. A. Hosseini, G. Cheung, and P. A. Chou, “Interpretable lightweight transformer via unrolling of learned graph smoothness priors,” NeurIPS, 2024.

AuthorsOpenReview ↗

Response (2) to Reviewer WMod

Features usedPCC
Signal only0.240
Signal + balanced0.278
Signal + frustrated0.416
Signal + balanced + frustrated0.439

Response to Weakness 3:

We thank the reviewer for this observation. We would like to first clarify the role of Proposition 1. Proposition 1 is intended as a theoretical motivation for adopting a near-balanced signed-graph model. Specifically, it shows that under a natural statistical construction, affinity graphs tend to exhibit a small number of frustrated edges, leading to near-balanced rather than fully balanced or highly frustrated structures. The proposition therefore motivates the graph prior used in the M-step.

We agree that Proposition 1 does not by itself explain downstream prediction performance, nor was it intended to do so. The performance benefit of the proposed near-balanced graph model arises from the resulting graph-learning and graph-filtering architecture and is evaluated empirically through the ablation study in Table 2. In particular, the fully balanced graph achieves denoising performance comparable to the near-balanced model while yielding substantially worse BDI prediction performance, indicating that preserving a small amount of frustration can improve the discriminative information retained for the downstream task.

We also agree that the graphs appearing in Proposition 1 and those learned in the M-step are not identical objects. Proposition 1 concerns thresholded affinity graphs and serves as theoretical justification for the near-balanced assumption, whereas the M-step learns task-adaptive near-balanced graphs from data. We will revise the manuscript to make this distinction and the intended role of Proposition 1 more explicit.

As for why the fully balanced graph achieves denoising MSE comparable to the near-balanced graph while resulting in substantially worse PCC, we have collected additional statistics showing that the balanced component accounts for 82.78%–83.82% of the edges across the three unrolled neural layers and 90.40%–91.45% of the total absolute edge-weight mass. Thus, recovering the balanced component alone is largely sufficient for signal reconstruction, explaining the similar denoising performance. However, given the learned near-balanced graph structures, we further evaluated the contribution of different graph components to depression-score prediction as shown in the provided table.

These results suggest that, while frustrated edges account for only a small fraction of the graph structure, they contain information that is highly relevant for discriminating between depressed and healthy subjects. In particular, the frustrated component alone accounts for the majority of the PCC gain, explaining why a fully balanced graph can achieve comparable denoising performance yet substantially inferior downstream prediction performance.

Response to Question 1:

We thank the reviewer for this question. If by "anti-correlated structure" the reviewer refers to information represented by negative edges and frustration, then our additional analysis provides direct evidence of its utility. Using the learned near-balanced graph structures, we evaluated downstream prediction performance using different graph components as shown in the provided table.

We observe that the frustrated component alone yields the majority of the PCC improvement, whereas the balanced component contributes only a modest gain. This suggests that information associated with frustration and signed interactions carries substantial discriminative information for the downstream depression-prediction task. Furthermore, Table 2 already shows that the fully-balanced graph achieves denoising MSE comparable to the near-balanced model while yielding substantially worse PCC. Taken together, these results indicate that retaining a small amount of frustration enables the model to preserve information that is largely suppressed when a strictly balanced graph is enforced. We emphasize that these observations concern the learned computational representation and attention structure of the unrolled transformer rather than biological brain connectivity. The learned signed graph is not intended to estimate anatomical or functional brain networks.

AuthorsOpenReview ↗

Response (3) to Reviewer WMod

Response to Question 2:

We thank the reviewer for this suggestion. To directly evaluate the necessity of the balanced/frustrated decomposition, we conducted an ablation study using the same learned near-balanced graph Laplacian Lθ\mathcal{L}_\theta​ and the same three-layer unrolled architecture. If the decomposition were unnecessary, then directly filtering with the full signed Laplacian should perform comparably to the proposed split-filtering strategy under the same finite-depth computation budget.

We compared three configurations:

  • Joint: filtering using the full near-balanced Laplacian Lθ\mathcal{L}_\theta​.
  • Split (CG/CG): filtering LθB\mathcal{L}_\theta^B​ and Fθ\mathbf{F}_\theta​ separately using CG for both components.
  • Split (Lanczos/CG): the proposed approach, using Lanczos approximation for LθB\mathcal{L}_\theta^B​ and CG for Fθ\mathbf{F}_\theta​. | Filter configuration | Reconstruction MSE ↓ | |:---|---:| | Joint | 0.14387786 | | Split (CG/CG) | 0.14395270 | | Split (Lanczos/CG) | 0.14215672 |

The proposed split (Lanczos/CG) strategy achieves the lowest reconstruction MSE among all tested configurations. This is consistent with the theoretical discussion in Appendices G and H. Specifically, the balanced component LθB\mathcal{L}_\theta^B​ corresponds to a positive graph whose low-frequency eigenspace captures the dominant smooth signal content. A finite-depth Lanczos approximation can therefore target this subspace efficiently, whereas directly filtering the full signed Laplacian cannot exploit this structure as effectively.

We stress that the advantage arises because the denoising process is implemented as a finite-depth unrolled network, rather than by solving (I+μLθ)x∗=y(\mathbf{I} + \mu \mathcal{L}_\theta) \mathbf{x}^*= \mathbf{y} to full convergence. In the convergence limit, the different formulations approach the same solution. Under finite unrolling, however, the decomposition enables a more effective parameterization of the filtering process, leading to improved signal reconstruction performance.

Response to Question 3:

We thank the reviewer for this question. The proposed model is intentionally lightweight. In the audited LOSO implementation, the denoiser contains 1,932 trainable parameters. The downstream linear BDI predictor adds 9 coefficients for the 8-dimensional signal representation (including intercept), plus 12 additional coefficients when graph-correction features are included. Thus, the complete denoiser-plus-predictor pipeline contains only 1,941 parameters (signal-only predictor) or 1,953 parameters (signal + graph-correction predictor) per fold. The parameter breakdown is:

ComponentParameters
Denoiser1,932
Signal readout9
Graph correction12
Total (signal only)1,941
Total (signal + graph correction)1,953

Following standard practice, PCA loadings and centering statistics are treated as fitted preprocessing transforms and are counted separately from trainable model parameters. We will clarify this counting convention and report full parameter counts in the revised manuscript for transparency.

Response to Question 4:

We thank the reviewer for this question. This concern is closely related to Weakness 3, and our response there addresses both issues. Briefly, Proposition 1 is intended as a theoretical motivation for the near-balanced graph prior used in the M-step. It shows that naturally occurring affinity graphs tend to be near-balanced, thereby motivating the near-balanced assumption. It is not intended to directly predict downstream task performance. The learned graphs in the M-step are task-adaptive graphs constructed from learned features, whereas Proposition 1 concerns thresholded affinity graphs; we will revise the manuscript to make this distinction more explicit.

Regarding the negative PCC of the fully-balanced graph, our additional analysis shows that the balanced component accounts for more than 90% of the total absolute edge-weight mass, making it largely sufficient for signal reconstruction and denoising. However, experiments using different graph components for prediction reveal that the frustrated component contributes the majority of the PCC improvement. Thus, a fully-balanced graph can achieve denoising MSE comparable to the near-balanced model while discarding information that is highly informative for the downstream depression-prediction task.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of our responses

We thank the Area Chair and reviewers for their thoughtful feedback. We would like to clarify three key points that underlie most of the raised concerns.

Interpretable unrolled transformer, not brain-network recovery. Our central contribution is an extremely parameter-efficient (<5% of SOTA model sizes) transformer-like neural network obtained by unrolling a near-balanced signed-graph optimization algorithm. The manuscript uses the term “interpretable” exactly once, in the standard algorithm-unrolling sense, where each neural layer corresponds to an iteration of a mathematically defined optimization procedure. The learned graphs are not intended to represent anatomical or functional brain connectivity; rather, as discussed in Appendix C, they function as layer-wise SELF-ATTENTION structures within the unrolled network.

Why near-balance and balanced/frustrated decomposition? Proposition 1 motivates the near-balanced graph prior but is not intended to predict downstream performance directly. Empirically, retaining a small amount of frustration is essential: while the fully balanced and near-balanced models achieve comparable denoising MSE, the PCC improves dramatically from -0.379 to 0.439. Additional analysis shows that the balanced component dominates signal reconstruction, whereas the frustrated component contributes disproportionately to downstream prediction performance.

Why split filtering if the exact solution is unchanged? We agree that the joint and split formulations share the same asymptotic denoising solution. However, our method unrolls a finite number of learned neural layers rather than solving the optimization problem to convergence. In this finite-depth setting, the parameterization matters. Direct filtering of the full signed Laplacian and split CG/CG filtering achieve nearly identical MSEs (0.14388 vs. 0.14395), while the proposed split Lanczos/CG implementation achieves the lowest MSE (0.14216), demonstrating the benefit of the decomposition for finite-step learned filtering.

Finally, we will clarify the alternating MAP/GEM interpretation, the role of Proposition 1, the LOSO training protocol, parameter counts, and the separation of pretraining, PCA, and downstream prediction stages in the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.