{"id":"3cdf4d31-6e68-4fbd-a798-e0258f0b0558","arxiv_id":"2412.02161","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Federated learning with LSTM and graph-attention models can predict node-level epidemic states on partitioned networks, with FedProx aggregation and balanced graph partitions performing best overall.","lead":"This paper tests whether federated learning, a technique that trains models on data kept in separate institutions without pooling it, can predict how epidemics spread across a network. On simulated outbreaks over a global airline network, the authors show that local temporal and graph-based models work reasonably well, but performance depends heavily on how the network is divided among clients and which aggregation rule is used.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"STGAT vs LSTM comparison is confounded by model capacity, so the claimed spatial-information advantage is not established.","rationale":"The reader's weakest assumption concerns the loss of cross-client edges during graph partitioning, which is a valid external-validity concern about whether spatial information survives in realistic deployments. My reading does not disagree that this is a risk, but the more immediate load-bearing gap is internal to the comparison: the paper's headline claim that STGAT excels because of graph-based spatial modeling is not supported by the experimental design, because STGAT differs from LSTM in capacity and depth as well as in graph usage. The partition issue would matter even if the comparison were clean, but the capacity confound must be resolved before any attribution to topology can be made. Since the paper is already CONDITIONAL and the proposed check is a feasible additional requirement rather than a reason to reject the whole empirical survey, I keep the reader's verdict unchanged while adding this specific condition. The check would settle the concern: if a capacity-matched LSTM matches STGAT, the central interpretational claim of the paper needs substantial revision.","tokens_in":23274,"tokens_out":8815,"duration_ms":98864,"concrete_test":"Train a two-layer LSTM baseline with 32 and 64 hidden units, matching STGAT's temporal stack and approximate parameter count, and evaluate it under the same centralized, FedAvg, and FedProx protocols on nmSIS, SIStv, and SIRS. If the matched LSTM's 1/CE and RMSE are statistically indistinguishable from STGAT's, the graph attention block is not the source of the reported advantage. As a complementary ablation, replace the GAT layer in STGAT with an identity or linear transformation while keeping all other components fixed; if performance does not drop, the claim that topology matters is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central conclusion that STGAT excels at capturing spatio-temporal dependencies rests on comparisons in which the models differ in more than graph usage. In Section 4.1.2, LSTM is specified with a single 64-hidden-unit LSTM layer, while STGAT consists of an 8-head graph attention layer followed by two LSTM layers with 32 and 64 hidden units. STGAT therefore has substantially more parameters and a deeper temporal stack. When STGAT outperforms LSTM on fluctuating dynamics such as nmSIS, SIStv, and SIRS, Section 4.1.3 attributes the gain to 'incorporation of graph-based attention mechanisms' and 'structural dependencies among nodes.' No capacity-matched LSTM baseline and no ablation of the GAT block is provided, so the observed advantage could be due to model size or temporal depth rather than topology. This is load-bearing because the paper's practical guidance--prefer STGAT for fluctuating dynamics and LSTM for simpler patterns--is meaningful only if the graph component itself contributes. The partition concern raised by the reader is related but secondary: even with a perfect partition that preserves all spatial information, the current experiments cannot separate the effect of graph structure from the effect of added capacity.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a federated learning framework for node-level epidemic prediction on networks, where the network is partitioned among clients that keep local data private and a central server aggregates model updates using FedAvg or FedProx. Two predictors are compared: a pure temporal LSTM and a spatio-temporal graph attention network (STGAT). The evaluation on a 600-node OpenFlights airline network covers seven epidemic models (SIS, SIR, SEIR, SIRS, SIR-VS, non-Markovian SIS, SIS with time-varying rates), varying numbers of clients, three graph partitioning schemes, different effective infection rates, and missing reporting. A summary statistic called 'efficacy energy' is introduced to aggregate performance across client counts. The central claims are that no single model dominates across all scenarios, FedProx generally outperforms FedAvg, and STGAT is preferable for fluctuating dynamics while LSTM suffices for simpler patterns.","tokens_in":23522,"tokens_out":4853,"duration_ms":46461,"significance":"If the results hold, the paper is a useful contribution to privacy-preserving epidemic forecasting: it formulates a realistic cross-silo setting, applies existing FL algorithms to a new domain, and proposes a compact metric (efficacy energy) for comparing FL configurations under variable client counts. The study explicitly acknowledges that STGAT does not universally beat LSTM, which is a more nuanced conclusion than typical 'graph-based model wins' narratives. The paper also ships source code on GitHub, which is helpful for reproducibility. However, the current experimental support is thin: a single stochastic realization, no error bars, absent epidemic parameter values, and a model comparison that confounds architecture capacity with graph usage all limit the reliability of the qualitative takeaways.","major_comments":[{"comment":"The conclusion that STGAT is superior for fluctuating dynamics (nmSIS, SIStv, SIRS) is not supported by the comparisons as run: the LSTM baseline is a single 64-unit LSTM layer, whereas STGAT adds an 8-head graph attention layer followed by two LSTM layers with 32 and 64 hidden units. The models therefore differ in parameter count and temporal depth, not only in graph usage. Without a capacity-matched LSTM or an ablation on the GAT block, the observed gains cannot be attributed to the graph-based attention mechanism, which is the load-bearing interpretation in Section 4.1.3 and Section 6.","section":"4.1.2–4.1.3"},{"comment":"The evaluation partitions the network by node index and trains each client on its induced subgraph, discarding all cross-client edges. When transmission across partition boundaries is significant, the local subgraph omits the very spatial interactions needed to predict boundary nodes, so the reported federated performance may be an artifact of the partition rather than a property of the FL framework. The paper should report the fraction of cut edges for each M and ideally include a condition in which the server shares topology (adjacency) with clients while keeping node states private, to separate privacy-preserving aggregation from the loss of global structural information.","section":"4.2.1"},{"comment":"Equation (20) defines the efficacy energy as η = 1/(M0-2) Σ_{K=2}^{M0} \\bar{α}[M], but the summation index K does not appear in the summand, and the denominator M0-2 does not match the M0-1 terms in the sum (from 2 to M0). This makes the metric not well-defined as written; it should be \\bar{α}[K] (or \\bar{α}[M]) with denominator M0-1. Since the paper's comparative conclusions are drawn from this aggregated quantity, this must be corrected.","section":"4.2.2, Eq. (20)"},{"comment":"No values are reported for any of the epidemic parameters listed in Table 1 (β, δ, ω, v1, v2, a, b, c, and the Weibull parameters for non-Markovian SIS), and the figures of prevalence (Figure 4) are not sufficient to determine them. In addition, the experiments use a single Gillespie realization per configuration with no error bars or multiple seeds. As a result, the claims that FedProx 'consistently outperforms' FedAvg and that STGAT outperforms LSTM in specific regimes are not supported by any measure of statistical significance or reproducibility.","section":"4.1.1–4.1.2, Table 1"}],"minor_comments":[{"comment":"The caption contains a typo: 'the greed for the recovered' should be 'the green for the recovered'.","section":"Figure 1 caption"},{"comment":"The text says 'they are used to forecast traffic speed in the next tF = 10 time steps'; this is a copy-paste error from a traffic-forecasting paper and should read 'forecast epidemic node states'.","section":"Section 4.1.2"},{"comment":"The phrase 'if Early Stop condition satisfried' should be 'satisfied'.","section":"Algorithm 1"},{"comment":"The caption and axis labels use 'missing radio' instead of 'missing ratio'.","section":"Figure 11"},{"comment":"The phrase 'allowing local users to access high-equality services' should be 'high-quality services'.","section":"Section 5.2"},{"comment":"Definition 2 allows 'disjoint or overlapping subnetworks', but the experiments only implement disjoint partitions; the paper should clarify that overlapping regions are out of scope for the numerical study.","section":"Section 3.1"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely and important problem, and the framework is reasonable, but the experimental evidence is not yet at the standard needed for the strong qualitative claims. The most serious issue is the STGAT vs LSTM capacity confound; the authors should either add a capacity-matched baseline and a GAT ablation, or substantially soften the causal attribution. The efficacy-energy formula error and the missing parameter values are straightforward to fix, but the lack of error bars and multiple seeds is a deeper issue that affects the credibility of the FedProx-vs-FedAvg comparison. I would be willing to reconsider after a major revision that addresses these points."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a reasonable, honest empirical study of federated learning for node-level epidemic prediction, but the main comparison it wants you to take away—STGAT wins on fluctuating dynamics, LSTM wins on simple ones—is not actually established. The two models differ in capacity and depth, not just in graph use. If the authors add a capacity-matched LSTM baseline or ablate the GAT block, the story may hold; right now it doesn't.\n\nWhat's genuinely here: a clear formulation of the FEP problem, coverage of seven epidemic models across Markovian and non-Markovian dynamics, a realistic 600-node airport network, and a systematic sweep over client count, aggregation (FedAvg vs FedProx), partition methods, and missing-report noise. The 'efficacy energy' is a simple average over client counts, but it is a reasonable summary statistic for robustness. The authors are honest that neither model dominates everywhere, and they explicitly flag the static-network limitation. Code is provided. That counts.\n\nSoft spots, in order of severity. First, the capacity confound: LSTM has one 64-unit layer; STGAT has an 8-head GAT plus two LSTM layers with 32 and 64 units. The gains attributed to 'graph-based attention' could equally come from extra parameters or the additional temporal layer. No capacity-matched baseline, no GAT ablation. That's the load-bearing issue. Second, the experiments appear single-run: no error bars, no seeds. For stochastic epidemic simulations and stochastic training, that matters. Third, the epidemic parameters (beta, delta, omega, v1/v2, a/b/c) are never reported; without them, others cannot reproduce or interpret the regime. Fourth, mechanical errors: Section 4.1.2 says the model forecasts 'traffic speed', Eq. (20) divides by M0-2 while summing M0-1 terms, and Figure 10's caption refers to nmSIS while the panels are SIS and SIRS. Those are fixable but need fixing.\n\nThe reader's worry about the node-index partition discarding cross-client edges is partially addressed because the paper later tests spectral and KL partitioning. That said, the default partition does throw away edges, so the STGAT advantage in the main analysis could be an artifact; the partition experiments help but don't rescue the capacity confound.\n\nVerdict: this deserves a serious referee, not a desk reject. A conditional accept after major revision is appropriate. The paper is for people working on applied federated learning for epidemic forecasting, not for someone seeking a fundamental advance.","headline":"Useful empirical sweep of federated epidemic prediction, but the STGAT-versus-LSTM comparison is confounded by model capacity and the evidence base lacks error bars; worth serious review with requested revisions.","tokens_in":24043,"tokens_out":2627,"would_cite":false,"duration_ms":26328,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that federated learning can deliver effective node-level epidemic forecasts on partitioned networks, with a spatio-temporal graph attention model (STGAT) preferred for fluctuating epidemic dynamics, LSTM sufficient for…","keywords":["federated learning","epidemic prediction","spatio-temporal graph attention networks","LSTM","efficacy energy","graph partitioning","network epidemiology","FedProx"],"falsifier":"Run the same federated pipeline on the same airline network and epidemic parameters, but partition nodes by geographic region or by a cut that removes the busiest inter-region routes, holding everything else fixed. If STGAT no longer outperforms LSTM, or if FedProx's advantage shrinks, then the reported results are an artifact of the chosen partition rather than a general property of federated epidemic prediction.","tokens_in":1611,"feed_emoji":"🦠","tokens_out":1798,"duration_ms":52077,"temperature":0.7,"pith_summary":"Federated epidemic prediction asks whether multiple data-holding regions can train a shared forecaster without pooling their health data. The paper argues the answer is yes for a broad family of network epidemic models — two-state and multi-state, Markovian and non-Markovian, static and time-varying infection rates — when each region trains locally and a central server aggregates model parameters. It shows that no single architecture wins everywhere: a spatio-temporal graph attention network (STGAT) is better at capturing fluctuating dynamics, while a plain temporal LSTM holds its own on simpler outbreak patterns. It also finds that FedProx aggregation consistently beats FedAvg as the number of clients grows, and introduces an \"efficacy energy\" metric to score a system's average performance across client configurations. The significance is practical: if these results hold, health authorities could collaborate on outbreak forecasting while keeping raw individual-level data local.","feed_headline":"Federated learning forecasts epidemics without pooling health data","feed_subtitle":"STGAT handles fluctuating outbreaks; FedProx beats FedAvg as client counts grow.","key_machinery":"The central mechanism is a federated training loop in which each client owns a subnetwork induced by a partition of the full network, trains locally on its own node-state trajectories, and shares only model parameters with a central server that aggregates them via FedAvg or FedProx. The local predictors are two architectures: a pure temporal LSTM and the proposed Spatio-Temporal Graph Attention Network (STGAT), which combines an embedding layer, a multi-head graph attention layer for spatial dependencies, two LSTM layers for temporal dependencies, and a softmax classifier. The paper's novel evaluation tool is \"efficacy energy,\" defined as the average of the mean client metric over an increasing number of clients, which measures system robustness under uncertain client configurations rather than at one fixed setting. Simulations use the Gillespie algorithm on a practical airline network, and graph partitions include even node-index splitting, spectral clustering, and Kernighan-Lin partitioning.","core_discovery":"On a 600-node OpenFlights airline network with epidemic trajectories generated by a Gillespie simulator, the paper claims that federated learning provides effective node-level epidemic prediction across seven compartmental models. In centralized baselines, LSTM performs better for SIS, SIR, SEIR, and SIR VS, while STGAT performs better for non-Markovian SIS, SIS with time-varying rates, and SIRS. In federated settings, FedProx consistently outperforms FedAvg, and its advantage grows with the number of clients, which the authors attribute to the proximal term stabilizing local updates under data heterogeneity. The paper also shows that graph partition strategy matters: spectral clustering helps when clients are few, Kernighan-Lin partitioning stays strong across client counts, and even index-based splits work reasonably. Performance degrades as the effective infection rate rises and as missing infectious reports increase, with the authors framing this as a tension between information richness and intrinsic stochasticity.","pith_inferences":["If the partition sensitivity seen in the paper is real, then real deployments where authorities hold only their own region's data may lose STGAT's advantage unless cross-client boundary information is shared through some privacy-preserving mechanism.","The efficacy energy metric is a reusable statistic for comparing any federated system under uncertain client participation, not just epidemic prediction, since it averages performance across client counts instead of relying on one configuration.","A direct testable extension would be to add synthetic cross-client transmission, such as shared boundary nodes or shared edge statistics, to measure how much of STGAT's gain comes from within-client topology versus global network structure.","The framework's logic could extend to other collective dynamics on networks, such as traffic congestion, cascading failures, or information spread, but the paper's static-network assumption leaves dynamic contact networks as an open challenge."],"forward_implications":["A central server can train node-level epidemic predictors across partitioned subnetworks without centralizing raw health data, enabling inter-region collaboration under privacy constraints.","FedProx should be preferred over FedAvg as the default aggregation method for federated epidemic prediction, because it degrades more gracefully as the number of clients increases.","For epidemic processes with fluctuating or periodic dynamics, the graph-aware STGAT model earns its added complexity, while for simpler outbreak patterns a temporal LSTM is sufficient.","Partition strategy is a first-order design choice: Kernighan-Lin partitioning balances feature consistency and volume uniformity across clients, while spectral clustering helps only when client counts are small.","Missing or noisy infectious reports degrade federated prediction performance, so deployments need data-quality safeguards such as noise-resilient aggregation or imputation."],"supporting_citations":[{"why":"Supplies the FedAvg aggregation rule that the paper uses as the baseline federated method.","marker":"[30]"},{"why":"Supplies the FedProx proximal-term aggregation that the paper finds more stable under data heterogeneity.","marker":"[31]"},{"why":"Provides the Gillespie-based event-driven simulator used to generate the epidemic trajectories on the airline network.","marker":"[24]"},{"why":"Defines the original STGAT architecture that the paper adapts for spatio-temporal epidemic prediction.","marker":"[17]"},{"why":"Defines the compartmental epidemic models and network epidemic framework that the experiments build on.","marker":"[2]"},{"why":"Provides the spectral clustering graph partition method compared in the partitioning experiments.","marker":"[34]"},{"why":"Provides the Kernighan-Lin graph partition method that shows consistently strong federated prediction performance.","marker":"[35]"}],"fun_headline_variants":["Federated LSTM and STGAT predict epidemics without sharing data","FedProx edges out FedAvg for federated outbreak forecasting","STGAT beats LSTM when disease dynamics are non-Markovian","Graph partition strategy shapes federated epidemic prediction","Federated learning forecasts outbreaks on real airline networks"],"cache_read_input_tokens":26112,"weakest_assumption_plain":"The central premise is that cutting the network into per-client subnetworks and discarding cross-client edges preserves enough spatial structure for the graph-based model to remain meaningful; if transmission across the cut edges dominates, the paper's STGAT results may not transfer to real deployments.","fun_headline_variants_meta":{"raw":{"variants":["Federated LSTM and STGAT predict epidemics without sharing data","FedProx edges out FedAvg for federated outbreak forecasting","STGAT beats LSTM when disease dynamics are non-Markovian","Graph partition strategy shapes federated epidemic prediction","Federated learning forecasts outbreaks on real airline networks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000686,"raw_usage":{"total_tokens":3135,"prompt_tokens":994,"completion_tokens":2141,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":2056}},"tokens_in":610,"tokens_out":2141,"duration_ms":14433,"temperature":1.0,"reasoning_tokens":2056,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:46:31.671379+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same federated pipeline on the same airline network and epidemic parameters, but partition nodes by geographic region or by a cut that removes the busiest inter-region routes, holding everything else fixed. If STGAT no longer outperforms LSTM, or if FedProx's advantage shrinks, then the reported results are an artifact of the chosen partition rather than a general property of federated epidemic prediction.","supporting_citations":[{"cited_title":"Federated optimization in heterogeneous networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the FedProx proximal-term aggregation that the paper finds more stable under data heterogeneity."},{"cited_title":"The spreading time in SIS epidemics on networks,","cited_arxiv_id":null,"evidence_quote":"Provides the Gillespie-based event-driven simulator used to generate the epidemic trajectories on the airline network."},{"cited_title":"Stgat: Modeling spatial-temporal interactions for human trajectory prediction,","cited_arxiv_id":null,"evidence_quote":"Defines the original STGAT architecture that the paper adapts for spatio-temporal epidemic prediction."},{"cited_title":"Epidemic processes in complex networks,","cited_arxiv_id":null,"evidence_quote":"Defines the compartmental epidemic models and network epidemic framework that the experiments build on."},{"cited_title":"Algebraic connectivity of graphs,","cited_arxiv_id":null,"evidence_quote":"Provides the spectral clustering graph partition method compared in the partitioning experiments."},{"cited_title":"An efficient heuristic procedure for partitioning graphs,","cited_arxiv_id":null,"evidence_quote":"Provides the Kernighan-Lin graph partition method that shows consistently strong federated prediction performance."}],"review_version":1}