{"id":"19b7abd8-9ffc-4caf-9275-b058b32c622e","arxiv_id":"2501.00442","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"SLoG-Net unrolls ADMM iterations into a trainable network that localizes sparse sources of graph diffusion with accuracy on par with the iterative solver and much faster inference.","lead":"The paper presents SLoG-Net, a neural network built by unrolling the ADMM optimization algorithm to locate sparse sources of graph diffusion when the diffusion filter is unknown. It aims to match a model-based solver's accuracy while cutting inference time by orders of magnitude.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ADMM baseline hyperparameters are not reported; 'on par' claim may be unfair without a tuned-baseline comparison.","rationale":"The reader identified Assumption 2 (filter invertibility) as the weakest assumption, but this is a stated scope limitation that holds almost surely in the synthetic experiments and does not directly threaten the central performance comparison. The more load-bearing concern is whether the ADMM baseline is represented fairly: the paper's main claim is that SLoG-Net is on par with ADMM while avoiding manual tuning, so the baseline's own hyperparameters must be at least reasonably selected. The absence of any reporting or tuning of ρλ and ρμ in the experiments leaves open the possibility that the accuracy gap is inflated. A concrete grid-search test would settle this. If the concern lands, the verdict should be CONDITIONAL pending a tuned-baseline comparison; the method itself may still be sound, but the central claim as stated needs support.","tokens_in":47482,"tokens_out":6607,"duration_ms":62529,"concrete_test":"Re-run the ADMM baseline of Fig. 4 and Table II with a grid search over ρλ and ρμ (e.g., log-spaced from 1e-3 to 1e3), selecting the best mean RE/ACC on the validation set for each test condition. If the best-tuned ADMM achieves RE/ACC statistically equal to or better than SLoG-Net at P=400, the 'on par' claim is only supported against an under-tuned baseline and should be revised; if SLoG-Net still matches or beats the best-tuned ADMM, the concern is resolved.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that SLoG-Net performs on par with the iterative ADMM baseline while being faster and avoiding manual tuning. The ADMM solver in Section III-B has penalty parameters ρλ and ρμ that strongly affect convergence and solution quality. However, the numerical evaluation in Section V gives no details on how these were chosen for the baseline (no grid search, no sensitivity analysis, no stopping criterion). If ADMM was run with default or arbitrary parameters, the accuracy comparison could be significantly handicapping the baseline, making the 'on par' claim overstate SLoG-Net's relative performance. This is load-bearing because the paper explicitly motivates SLoG-Net as avoiding the need to tune such parameters; it must therefore demonstrate that it matches a reasonably tuned ADMM.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies blind deconvolution for source localization on graphs: given diffused graph signals Y generated from sparse sources X through an unknown graph filter H, it jointly estimates the inverse filter and the sources. Under an invertibility assumption on H, the authors reformulate the problem as a convex ℓ1-synthesis program, derive an ADMM solver, unroll the iterations into a trainable network called SLoG-Net, and evaluate the method on synthetic and real data. The central empirical claim is that SLoG-Net achieves accuracy on par with the iterative ADMM baseline while being orders of magnitude faster at inference and avoiding manual tuning of step-size or penalty parameters.","tokens_in":47638,"tokens_out":8345,"duration_ms":91319,"significance":"If the comparison is fair and the empirical claims hold, this is a useful contribution to the algorithm-unrolling literature for graph inverse problems. The paper gives a clean derivation of the ADMM updates, exploits the diagonal structure of Z^T Z and the matrix inversion lemma to make the model-based solver practical, and provides reproducible code for the experiments. The architecture is interpretable and parameter-efficient by construction, and the noise-robustness experiments show graceful degradation. The real-data experiment is honestly described as difficult, although the reported effect sizes are small.","major_comments":[{"comment":"The ADMM baseline is not fully specified in the numerical evaluation. The penalty parameters ρλ and ρμ, the stopping criterion, the maximum number of iterations, and the initialization are never reported, and it is not stated whether the experiments use the original constraint 1^T g = c or the modified constraint M^T g = m from Algorithm 1. Since the paper motivates SLoG-Net by the need to avoid manually tuning exactly these parameters, the comparison must demonstrate that ADMM was reasonably tuned; otherwise the 'on par' accuracy claim and the timing comparison in Table I could reflect an arbitrarily poor or slow baseline rather than a genuine property of the unrolled network. Please report the grid search or sensitivity analysis for ρλ and ρμ, the convergence criterion and iteration cap, and the exact ADMM variant used.","section":"Section V-B with Section III-B"},{"comment":"The real-data experiment reports a mean AUC of 0.56 for SLoG-Net versus 0.51 for IVGD, which is close to chance, and no error bars, confidence intervals, or per-subgraph results are given. The sentence in Section V-E that SLoG-Net 'is better at learning representations that are predictive of the sources' is therefore stronger than the displayed evidence supports. Please provide the spread over the 10 realizations (and ideally a paired comparison), or temper the conclusion accordingly.","section":"Section V-E"}],"minor_comments":[{"comment":"Equation (16) defines the sources sub-layer with only α1 and α2, but the caption of Fig. 2 lists Θ_X^{(k)} = {α1, α2, α3, τ}. Please clarify what α3 is used for or remove it from the caption.","section":"Section IV-A and Fig. 2"},{"comment":"The hyperparameter selection paragraph says that K = 5 and d = 2 were chosen after a grid search, but the grid values, the validation criterion, and the cost of the search are not given; please include these details to support the reproducibility claim.","section":"Section V-A"},{"comment":"The pseudocode in Fig. 2 uses notation such as ωω, ωµ, and ε^{-1}ω that does not match the ρλ, ρμ, and ρλ^{-1} used in Section III-B, and several occurrences of '↑' appear where a subtraction sign is intended; this makes the algorithm hard to read and should be corrected.","section":"Fig. 2, Algorithm 1"},{"comment":"The wall-clock timing comparisons do not report the hardware, software versions, or random seeds used; please add these details so that the claimed inference speedups can be reproduced and compared fairly.","section":"Section V-B and Table I"},{"comment":"There are several typos that should be corrected in revision, including 'syntethic' in Section IV-A, 'corruptin' in Section V-B, 'mangnitude' in Section V-B, 'pr thousands' in Section IV, and 'consisent' in Section V-A.","section":"Entire text"},{"comment":"The text says initial states can be randomly initialized 'as we do in the ensuing experiments,' while Section III-B suggests zero initialization is typical; please state explicitly what initialization is used for the ADMM baseline and for SLoG-Net at both training and inference.","section":"Section IV-A and Section V-A"}],"recommendation":"major_revision","confidential_remarks":"The main issue is the fairness and completeness of the ADMM baseline comparison. I do not see a fundamental flaw in the derivation, and the concern is fixable by reporting and, ideally, supplementing the experiments with a tuned baseline. The real-data AUC is weak but the authors are candid about the difficulty; I would not reject on that basis alone."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read SLoG-Net. Bottom line: it does what it says. The ADMM derivation is clean, the unrolling is faithful, and the experiments back the accuracy claims. The main thing to fix before publication is the ADMM baseline: the penalty parameters are never specified, so the 'on par' claim is under-supported as reported.\n\nWhat's new: this is the first ADMM unrolling for blind graph deconvolution on graphs, as far as the cited literature goes. The architectural refinements—learnable constraints M, m, decoupled per-layer parameters—are real changes, not a trivial transfer. They also give a standalone ADMM solver for the convex relaxation (8), which is useful in its own right. The matrix inversion lemma trick in Appendix C is correct and gives an O(N^2) inversion, and the paper shares code.\n\nWhat's good: the derivations in Section III and Appendix A check out, the experiments are honest and cover noise, P, N, and several graph types with ten-run error bars. Limitations are disclosed: the invertibility assumption on H, small graph sizes, and the weak real-data AUC. None of these invalidates the stated claims.\n\nThe soft spot: the ADMM baseline's rho_lambda and rho_mu are not reported, and there's no grid search or sensitivity analysis. Those penalty parameters materially affect how fast ADMM converges and what solution it stops at. Since the paper's pitch is 'no manual tuning,' the comparison has to be against a reasonably tuned ADMM. As written, the timings and the 'on par' accuracy could be an artifact of a badly-tuned baseline. This is load-bearing but fixable: report the parameters, show a sensitivity sweep, and ideally include a tuned baseline.\n\nMinor: the real-data AUC (0.56 vs 0.51) is weak and they say so; the synthetic graphs cap at 100 nodes, so scalability is modest. The invertibility assumption is a scope condition, not a hidden flaw.\n\nCitation pattern is appropriate. Self-citation of [42] is justified because the convex relaxation is the foundation; the new result is the unrolling and the solver. No circularity.\n\nWho it's for: GSP researchers, people working on algorithm unrolling, and anyone doing network inverse problems. It's a solid, honest paper and deserves a serious referee. I'd send it out and ask for the ADMM baseline details as a major revision item.","headline":"A clean ADMM-unrolling paper with an honest comparison; just pin down the ADMM baseline's penalty parameters before accepting the 'on par' claim.","tokens_in":48119,"tokens_out":2665,"would_cite":true,"duration_ms":25173,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"By unrolling ADMM iterations, SLoG-Net turns blind source localization on graphs into a fast trainable network that matches the iterative solver's accuracy.","keywords":["graph signal processing","source localization","blind deconvolution","algorithm unrolling","ADMM","graph filters","sparse recovery","network diffusion"],"falsifier":"Take a graph and a filter $H$ whose frequency response is exactly zero at one graph eigenvalue, generate $Y=HX$ from sparse sources $X$, and run both the convex program (8) and a trained SLoG-Net on the observations; if either method still recovers the source support, then the invertibility assumption is not necessary as claimed, and if neither recovers it, the assumption is doing the load-bearing work.","tokens_in":47303,"feed_emoji":"🕸️","tokens_out":7957,"duration_ms":70031,"temperature":0.7,"pith_summary":"The paper claims that locating the source nodes of a network diffusion process, when the diffusion filter is unknown, reduces to a convex ℓ1-synthesis problem once the filter is invertible, and that unrolling the ADMM solver for that problem yields a neural network that matches the iterative baseline's accuracy while running orders of magnitude faster. The resulting model, SLoG-Net, learns step sizes, penalty parameters, thresholds, and the scale constraint from labeled data, removing the manual tuning that the ADMM baseline needs. The authors support the claim with experiments on synthetic random graphs, a community-detection comparison with a graph neural network, and a real-data study on Digg voting cascades.","feed_headline":"Unrolled ADMM finds network diffusion sources at ADMM accuracy","feed_subtitle":"SLoG-Net turns ADMM updates for blind graph deconvolution into a trainable network, matching accuracy at a fraction of the inference time.","key_machinery":"The argument is carried by the invertible-graph-filter assumption and the frequency-domain identity $\\tilde{y} = \\operatorname{diag}(\\Psi_L h)\\tilde{x}$. Under invertibility, the inverse filter $G=H^{-1}$ exists as a polynomial in the graph-shift operator, so $X=GY$ replaces the bilinear forward model with a linear one; the sparsity objective $\\|X\\|_{1,1}$ then factors as $\\|(Y^\\top V \\odot V)\\tilde{g}\\|_1$, a convex linear program in the inverse filter's frequency response. ADMM solves this program with a filter update that inverts a rank-one correction of a diagonal matrix, a soft-thresholding source update, and dual multiplier updates; SLoG-Net unrolls these three steps into corresponding sub-layers with learnable combination weights, thresholds, penalty parameters, and constraint vectors $M$ and $m$, which is what lets the network be trained end-to-end from labeled data.","core_discovery":"On the paper's own terms, the central discovery is that source localization on graphs can be solved through a convex program and that the ADMM iterations of that program can be truncated, parameterized, and trained without losing accuracy. With observations modeled as $Y=HX$ for a polynomial graph filter $H$, the invertibility of $H$ lets the authors rewrite the bilinear model as $X=GY$, where $G=H^{-1}$ is also a graph filter; estimating the inverse filter's frequency response $\\tilde{g}$ then becomes the linear program $\\min \\|(Y^\\top V \\odot V)\\tilde{g}\\|_1$ subject to $1^\\top \\tilde{g} = 1$. The paper derives ADMM updates for this program, exploits the diagonal structure of $Z^\\top Z$ to invert the needed matrix in $O(N^2)$ time, and unrolls the updates into a $K$-layer network with learnable parameters in each filter, sources, and multipliers sub-layer. The numerical results show SLoG-Net on par with ADMM on recovery error and support accuracy, with better noise robustness and inference times around $10^{-2}$ seconds compared with seconds for ADMM.","pith_inferences":["Because SLoG-Net is trained to imitate the convex relaxation (8), its performance ceiling is likely the convex program's recovery region; a testable extension is to train against the non-convex $\\ell_0$ objective or filters outside the training distribution to see whether unrolling can surpass that baseline.","The architecture drops the common-support assumption of earlier lifted-matrix methods, so a natural next test is whether the layer-wise decoupled parameters transfer to larger graphs than those seen in training, which the paper lists as future work.","The invertibility assumption suggests an experimental design rule: choose diffusion dynamics whose frequency response stays bounded away from zero on the graph spectrum, because spectral nulls make both the convex program and its unrolled version blind to the lost modes.","The binary-data calibration trick, which chains an invertible residual network in front of SLoG-Net, indicates a broader pattern where unrolled graph-signal networks can be combined with learnable pre-processors whenever observations violate the diffusion model."],"forward_implications":["SLoG-Net matches the ADMM baseline on source recovery error and support accuracy while running roughly two to three orders of magnitude faster at inference, so model-based convex relaxations can be converted into deployable inference engines.","Because step sizes, penalty coefficients, thresholds, and the scale constraint are learned, users no longer need to hand-tune those quantities for each graph or noise level.","The architecture stays interpretable: each layer corresponds to an ADMM update, so behavior can be inspected and complexity can be controlled by choosing the number of layers $K$.","On the Digg 2009 data, SLoG-Net achieves a higher AUC than the IVGD baseline in identifying which users' early votes sparked a story's voting cascade, indicating the approach transfers beyond synthetic diffusion models.","The same ADMM blueprint applies to other blind graph deconvolution tasks that satisfy the invertibility condition, since the network is built from the solver updates rather than from task-specific features."],"supporting_citations":[{"why":"supplies the convex $\\ell_1$-synthesis relaxation (8) and the exact-recovery and noise-stability guarantees that both the ADMM solver and SLoG-Net are built on.","marker":"[42]"},{"why":"establishes identifiability for the multi-signal blind deconvolution problem with sparse Bernoulli-Gaussian inputs, motivating the supervised formulation and the invertible-filter maneuver.","marker":"[44]"},{"why":"provides the earlier blind graph-filter identification formulation and the matrix-lifting viewpoint that this paper avoids for computational reasons.","marker":"[33]"},{"why":"supplies the algorithm-unrolling methodology that turns truncated ADMM iterations into the SLoG-Net architecture.","marker":"[23]"},{"why":"is the ADMM-unrolling precedent for image compressive sensing that the authors adapt to graph blind deconvolution.","marker":"[40]"},{"why":"is the IVGD baseline used in the real-data experiment and the source of the invertible-residual calibration idea for binary observations.","marker":"[36]"},{"why":"provides the Digg 2009 dataset of voting cascades and friendship links used in the real-data source-localization test.","marker":"[14]"}],"fun_headline_variants":["SLoG-Net: trainable unrolled ADMM for graph source localization","Unrolling ADMM into a neural net: same accuracy, faster inference","Graph source localization: from convex ADMM to learned unrolling","SLoG-Net: deep unrolling of ADMM for fast graph source finding"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the diffusion filter $H$ is invertible, meaning its frequency response is nonzero at every graph eigenvalue, so that the observations can be rewritten as $X=GY$ with $G=H^{-1}$ and the convex $\\ell_1$ program applies.","fun_headline_variants_meta":{"raw":{"variants":["SLoG-Net: trainable unrolled ADMM for graph source localization","Unrolling ADMM into a neural net: same accuracy, faster inference","Graph source localization: from convex ADMM to learned unrolling","SLoG-Net: deep unrolling of ADMM for fast graph source finding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000474,"raw_usage":{"total_tokens":2399,"prompt_tokens":1033,"completion_tokens":1366,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":649,"completion_tokens_details":{"reasoning_tokens":1284}},"tokens_in":649,"tokens_out":1366,"duration_ms":12088,"temperature":1.0,"reasoning_tokens":1284,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:50:37.148232+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a graph and a filter $H$ whose frequency response is exactly zero at one graph eigenvalue, generate $Y=HX$ from sparse sources $X$, and run both the convex program (8) and a trained SLoG-Net on the observations; if either method still recovers the source support, then the invertibility assumption is not necessary as claimed, and if neither recovers it, the assumption is doing the load-bearing work.","supporting_citations":[{"cited_title":"Blind deconvolution on graphs: Exact and stable recovery,","cited_arxiv_id":null,"evidence_quote":"supplies the convex $\\ell_1$-synthesis relaxation (8) and the exact-recovery and noise-stability guarantees that both the ADMM solver and SLoG-Net are built on."},{"cited_title":"Blind identification of invertible graph filters with multiple sparse inputs,","cited_arxiv_id":null,"evidence_quote":"establishes identifiability for the multi-signal blind deconvolution problem with sparse Bernoulli-Gaussian inputs, motivating the supervised formulation and the invertible-filter maneuver."},{"cited_title":"Blind identifi- cation of graph filters,","cited_arxiv_id":null,"evidence_quote":"provides the earlier blind graph-filter identification formulation and the matrix-lifting viewpoint that this paper avoids for computational reasons."},{"cited_title":"ADMM-CSNet: A deep learning approach for image compressive sensing,","cited_arxiv_id":null,"evidence_quote":"is the ADMM-unrolling precedent for image compressive sensing that the authors adapt to graph blind deconvolution."},{"cited_title":"An invertible graph diffusion neural network for source localization,","cited_arxiv_id":null,"evidence_quote":"is the IVGD baseline used in the real-data experiment and the source of the invertible-residual calibration idea for binary observations."},{"cited_title":"Social dynamics of Digg,","cited_arxiv_id":null,"evidence_quote":"provides the Digg 2009 dataset of voting cascades and friendship links used in the real-data source-localization test."}],"review_version":1}