{"id":"861aa8db-aad8-4853-8a3d-f345f6b75ef1","arxiv_id":"2512.01708","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":7.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Fed-Sparse-BNSL combines differential privacy with sparse greedy updates to learn linear Gaussian Bayesian network structures in a federated setting while keeping communication low and utility close to non-private baselines.","lead":"Fed-Sparse-BNSL is a federated algorithm that learns linear Gaussian Bayesian network structures from decentralized data by adding differential privacy noise and restricting communication to a few relevant edges per participant. This could let organizations build joint probabilistic models on sensitive distributed data without sharing raw records or incurring high communication costs.","discovery_kind":"unclear","skeptic_critique":{"model":"grok-4.3","headline":"DP noise on sparse per-participant edge updates may systematically bias score-based selection away from the true linear-Gaussian structure","rationale":"The reader’s weakest assumption directly flags the same point. Because the full manuscript is now available, the concrete test above can be executed on the reported experimental setup; a negative result would move the verdict from UNVERDICTED to CONDITIONAL while a positive result would support keeping the current claim.","tokens_in":1618,"tokens_out":369,"duration_ms":22488,"concrete_test":"Re-run the synthetic experiments of §5.1 on the 10-node chain and 20-node random DAG using the exact privacy parameters reported in the paper; compute structural Hamming distance and edge-precision for both Fed-Sparse-BNSL and the non-private baseline. If the median SHD increases by >15 % or precision drops below 0.85 while the non-private version stays above 0.95, the bias concern is confirmed.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim requires that restricting each participant to a few candidate edges and releasing noisy scores (or gradients) still permits consistent recovery of the global DAG. In linear-Gaussian score-based learning the local scores are typically BIC or marginal likelihood terms whose differences can be small; adding calibrated DP noise (Gaussian or Laplace) perturbs the argmax ordering. If the noise scale required for the stated ε,δ exceeds the typical score gap on the true versus false edges, the greedy selection step can prefer spurious parents, violating both identifiability and the “utility close to non-private” guarantee. The paper’s algorithmic design claims to preserve identifiability, yet no explicit bound relating noise variance, score sensitivity, and minimum edge strength appears to be derived.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript proposes Fed-Sparse-BNSL, a federated algorithm for structure learning of linear Gaussian Bayesian networks. It integrates differential privacy with per-participant greedy updates restricted to a small number of candidate edges to control communication cost and privacy budget. The central claims are that the algorithmic design preserves model identifiability and that experiments on synthetic and real datasets show utility close to non-private baselines while providing stronger privacy and efficiency.","tokens_in":1789,"tokens_out":495,"duration_ms":35593,"significance":"If the identifiability claim and near-baseline utility hold, the work would be a useful contribution to privacy-preserving federated graphical model learning, especially for high-dimensional settings where full communication is prohibitive. The sparse-update strategy combined with DP is a pragmatic design choice that could enable deployment in domains with decentralized sensitive data.","major_comments":[{"comment":"The assertion that 'careful algorithmic design preserves model identifiability' is load-bearing for the central claim yet lacks an explicit bound relating DP noise scale (Gaussian or Laplace), score sensitivity, and minimum edge strength. In linear-Gaussian score-based learning, BIC or marginal-likelihood differences between true and spurious parents can be small; without a derived condition ensuring that the calibrated noise does not flip the argmax ordering in the greedy step, consistent recovery of the true DAG is not guaranteed.","section":"Algorithmic Design / Theoretical Analysis"},{"comment":"The experimental claim of 'utility close to non-private baselines' is not yet supported by evidence that the chosen privacy parameters keep noise below typical score gaps on the evaluated datasets. If the noise variance required for the stated ε,δ exceeds the separation between correct and incorrect edges, the reported performance may be confined to favorable regimes and does not substantiate the general utility guarantee.","section":"Experiments"}],"minor_comments":[{"comment":"The abstract states 'substantially stronger privacy' without quoting the concrete (ε,δ) values or the privacy budget allocation across rounds; adding these numbers would improve clarity.","section":"Abstract"},{"comment":"Notation for the local score function and the sensitivity bound used in the DP mechanism should be introduced earlier and used consistently when describing the noisy update step.","section":"Methods"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive and detailed comments on our manuscript. We address each major comment below and outline the revisions we intend to make to strengthen the presentation of our results.","responses":[{"response":"We acknowledge that the current manuscript does not supply an explicit theoretical bound relating the DP noise scale to score sensitivity and minimum edge strength. Our identifiability claim rests on the observation that restricting each participant to a small set of candidate edges substantially reduces the sensitivity of the local score computations, thereby limiting the impact of the calibrated Gaussian noise on the greedy argmax step. We agree that a more formal condition would be valuable. In the revision we will expand the theoretical analysis section with a discussion of score sensitivity for linear-Gaussian models, explain how the sparse-update restriction helps preserve ordering under the added noise, and state the additional assumptions (e.g., minimum edge strength and data distribution) that would be needed for a worst-case guarantee. We will not claim a complete proof if one is not derived.","revision_made":"partial","referee_comment":"[Algorithmic Design / Theoretical Analysis] The assertion that 'careful algorithmic design preserves model identifiability' is load-bearing for the central claim yet lacks an explicit bound relating DP noise scale (Gaussian or Laplace), score sensitivity, and minimum edge strength. In linear-Gaussian score-based learning, BIC or marginal-likelihood differences between true and spurious parents can be small; without a derived condition ensuring that the calibrated noise does not flip the argmax ordering in the greedy step, consistent recovery of the true DAG is not guaranteed."},{"response":"We appreciate this observation. To substantiate the utility claim, the revised experimental section will include a direct comparison, for each dataset and privacy parameter setting, of the noise variance induced by the chosen (ε, δ) against the observed score gaps (BIC or marginal-likelihood differences) between correct and incorrect parent sets. This analysis will be presented alongside the existing performance tables so that readers can verify that the noise remains below the typical separation in the evaluated regimes.","revision_made":"yes","referee_comment":"[Experiments] The experimental claim of 'utility close to non-private baselines' is not yet supported by evidence that the chosen privacy parameters keep noise below typical score gaps on the evaluated datasets. If the noise variance required for the stated ε,δ exceeds the separation between correct and incorrect edges, the reported performance may be confined to favorable regimes and does not substantiate the general utility guarantee."}],"tokens_in":1292,"tokens_out":531,"duration_ms":36184,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"This paper's main contribution is Fed-Sparse-BNSL, which does federated structure learning for linear Gaussian Bayesian networks under differential privacy by limiting each participant to updating only a few candidate edges. That keeps both the privacy budget and the communication costs manageable. The idea of sparse per-participant greedy updates combined with DP noise addition seems tailored to avoid the usual scaling issues in high dimensions. They show on synthetic and real datasets that the utility stays close to the non-private case while adding privacy and cutting comms. The algorithmic choices around sparse updates and DP mechanisms look like a sensible engineering response to the decentralized setting, and the experiments back up the practical claims at least for the tested regimes. The soft spot is the lack of an explicit analysis showing that the added noise does not systematically distort the greedy edge selection. In score-based learning the margins between correct and incorrect parents are often modest, so the noise level needed for meaningful privacy could flip some decisions. The claim that identifiability is preserved would be stronger with a bound relating noise scale, score sensitivity, and minimum signal strength; right now it seems to rest on the experimental outcomes. If the full paper has such a derivation I missed it in the abstract, but the stress test concern about bias in argmax ordering looks worth checking. A reader working on privacy-aware graphical models or federated structure learning would find the concrete method and the reported trade-offs useful. It is worth sending to peer review because the problem is practical and the approach is new enough to merit referee input on both the theory and the experiments.","headline":"The paper gives a workable federated DP algorithm for Bayesian network structure learning via sparse updates, though the noise may still bias edge selection without tighter guarantees.","tokens_in":2262,"tokens_out":387,"would_cite":false,"duration_ms":37699,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":{"model":"grok-4.3","evidence":[{"relation":"unclear","rs_module":"IndisputableMonolith/Foundation/RealityFromDistinction.lean","rs_theorem":"reality_from_one_distinction","paper_passage":"We introduce Fed-Sparse-BNSL, a novel federated method for learning linear Gaussian Bayesian network structures... combining differential privacy with greedy updates that target only a few relevant edges per participant"}],"headline":"Federated DP Bayesian network structure learning paper is orthogonal to RS forcing chain","alignment":"orthogonal","rationale":"The paper's central machinery (PGCD, DP-PGCD with Gumbel/Gaussian noise, ADMM consensus for linear-Gaussian BNs, sparse edge updates) addresses practical federated optimization and privacy accounting in high-dimensional DAG recovery. This has no overlap with RS concepts such as the reciprocal cost J, φ-ladder, 8-tick periodicity, or the reality_from_one_distinction forcing chain. The domain (stat.ML structure learning) is one on which RS has no opinion.","tokens_in":56577,"confidence":"high","tokens_out":240,"duration_ms":12174,"cache_read_input_tokens":38528,"cache_creation_input_tokens":0},"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A federated method learns linear Gaussian Bayesian network structures with differential privacy by restricting each participant to updating only a few relevant edges.","keywords":["Bayesian networks","structure learning","differential privacy","federated learning","linear Gaussian models","privacy-preserving machine learning","decentralized data analysis"],"falsifier":"A dataset where the non-private baseline recovers the correct edges but the private federated version consistently selects a substantially different edge set or fails to identify the true parents of key variables.","tokens_in":2532,"feed_emoji":"🔒","tokens_out":656,"duration_ms":76663,"temperature":0.7,"pith_summary":"The paper introduces Fed-Sparse-BNSL to learn Bayesian network structures from decentralized data without centralizing raw observations. It tackles privacy and communication scaling by combining differential privacy with greedy updates that touch only a small number of edges per participant. The design choices aim to keep the underlying linear Gaussian model identifiable so the true dependencies remain recoverable despite the added noise. Experiments on synthetic and real datasets show that the resulting structures achieve accuracy close to non-private centralized baselines while using far less communication and providing formal privacy guarantees.","feed_headline":"Sparse private updates recover Bayesian network structures","feed_subtitle":"Fed-Sparse-BNSL matches non-private accuracy on synthetic and real data while adding differential privacy and keeping communication low.","key_machinery":"Fed-Sparse-BNSL, which applies differential privacy to greedy updates that select and modify only a few relevant edges per participant.","core_discovery":"Fed-Sparse-BNSL performs federated structure learning for linear Gaussian Bayesian networks by having each participant apply differential privacy to a sparse set of greedy edge updates rather than transmitting full local statistics. Careful restriction of the updates to a few relevant edges per round preserves model identifiability, allowing accurate recovery of the global network structure without participants sharing their full datasets or incurring communication costs that grow with dimensionality.","pith_inferences":["The sparse-update idea could transfer to structure learning for other graphical models where full edge sets are costly to communicate.","Institutions holding sensitive data might use similar techniques to build joint models without pooling raw records.","Varying the number of edges updated per round offers a tunable knob for trading privacy strength against estimation accuracy.","The approach might extend to discrete or mixed-variable Bayesian networks if the identifiability arguments are adapted accordingly."],"forward_implications":["Communication cost stays low even as the number of variables grows because only a bounded number of edges are exchanged per participant.","Differential privacy can be added without destroying the ability to identify the correct network dependencies.","The method yields structures whose utility approaches that of non-private methods on both synthetic and real data.","Privacy budget is used efficiently by concentrating noise on the selected sparse updates rather than on all possible edges."],"fun_headline_variants":["Sparse private updates enable federated Bayesian structure learning","Fed-Sparse-BNSL learns Bayesian networks privately with low communication","Differentially private sparse edge updates for Bayesian network learning","Accurate Bayesian structures recovered via private federated sparse updates"],"cache_read_input_tokens":2496,"weakest_assumption_plain":"That restricting updates to a few relevant edges and adding differential privacy noise still permits accurate recovery of the true linear Gaussian Bayesian network structure without systematic bias or loss of identifiability.","fun_headline_variants_meta":{"raw":{"variants":["Sparse private updates enable federated Bayesian structure learning","Fed-Sparse-BNSL learns Bayesian networks privately with low communication","Differentially private sparse edge updates for Bayesian network learning","Accurate Bayesian structures recovered via private federated sparse updates"]},"model":"grok-4.3","cost_usd":0.008611,"raw_usage":{"total_tokens":3837,"prompt_tokens":570,"num_sources_used":0,"completion_tokens":50,"cost_in_usd_ticks":86112000,"prompt_tokens_details":{"text_tokens":570,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":3217,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":570,"tokens_out":50,"duration_ms":32989,"temperature":1.0,"reasoning_tokens":3217,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-05-17T02:44:31.509122+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A dataset where the non-private baseline recovers the correct edges but the private federated version consistently selects a substantially different edge set or fails to identify the true parents of key variables.","supporting_citations":[],"review_version":1}