{"id":"00bce396-bb02-4f6d-8189-08d3222178b4","arxiv_id":"2501.07327","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The LETN method adds node labels to egocentric temporal network signatures, producing surrogate temporal networks that better reproduce community-level interaction patterns than the unlabeled ETN-gen.","lead":"This paper extends a prior temporal-network generator, ETN-gen, by adding node labels that encode community membership or role, producing synthetic face-to-face interaction networks whose community structure is closer to the original. A generalist should care because realistic surrogate contact networks could support epidemic and opinion-dynamics simulations without collecting new private interaction data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The main validation metrics are label-conditioned statistics that the generator is trained to reproduce, so the headline claim of fidelity to community interaction structure is not independently established.","rationale":"The reader's weakest_assumption precisely identifies the circularity: the label-conditioned statistics used as validation are the same statistics encoded in the generator's probability dictionaries. My attack formalizes this by pointing to Section 2.1, where the LETNS encodes node labels, and the probability dictionaries are built from masked LETNS, so the transition distributions are explicitly conditioned on the labels of interacting nodes. Modularity and label assortativity measured with the original labels are therefore direct aggregates of the label-pair transition statistics that the generator is trained to reproduce. The comparison against ETN-gen is not a strong control because ETN-gen lacks labels entirely and is expected to fail these metrics. The paper does offer some independent evidence: the PCA visualization (Figure 3) and the interaction-duration heatmaps (Figure 5) are not direct aggregates of the label-pair transition probabilities and do support the claim that LETN captures more than raw label mixing. The CLETN/DLETN comparisons provide some validation that the label assignment strategy is robust, but they rely on the same circular metrics. The hold-out temporal test I propose is a concrete way to break the circularity, and it is one the authors could run with their existing data and code. Given the moderate confidence and the circularity concern, conditional acceptance is the appropriate verdict. Agreement with the reader is 'agree' because the reader's weakest_assumption is the same load-bearing concern I identify.","tokens_in":16436,"tokens_out":1633,"duration_ms":14471,"concrete_test":"Perform a hold-out temporal validation on one dataset (e.g., Hospital or High School 2013): build the LETN probability dictionaries from the first 60% of temporal layers, then generate surrogates and evaluate modularity, label assortativity, degree distribution, and interaction-duration distributions on the held-out 40% of layers. If the match degrades substantially compared to the in-sample match, the method is fitting the training statistics rather than capturing a stable generative mechanism. This test is targeted because it directly breaks the circular overlap between the training objective and the evaluation metrics.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's strongest evidence is the match in modularity and label assortativity between LETN-generated networks and the original datasets (Figures 6-7, Table 2). However, these metrics are direct aggregates of the exact label interactions the method encodes. In Section 2.1, LETNS includes the ego label and neighbor labels, and the probability dictionaries are built from masked LETNS, so the transition distributions are explicitly conditioned on the labels of interacting nodes. Modularity with the original labels and label assortativity measure precisely those label-conditioned mixing patterns. This creates a circularity: the generator is fit to reproduce label-pair transition statistics, and then evaluated on summary statistics of those same label-pair statistics. The ETN baseline, which lacks labels, is expected to fail these metrics, so the comparison is not a disconfirmation of the method's core mechanism. The paper does not hold out part of the temporal sequence, does not evaluate downstream dynamics (e.g., epidemic spread or information diffusion), and does not compare against an attributed generator baseline such as DyANE, which directly generates attributed networks. The PCA visualization and interaction-duration analysis are more independent evidence, but the main quantitative claim rests on the circular metrics. The limitation is not that the method is useless, but that the claimed validation of 'replicating realistic behaviors' is not yet demonstrated beyond reproducing the statistics it was designed to capture. This warrants conditional acceptance rather than full acceptance, pending a test that does not share the training objective. Please note that this assessment matches the reader's concern.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes Labeled ETN-gen (LETN), an extension of the ETN-gen surrogate temporal network generator. The method adds node labels (from metadata or community detection) to egocentric temporal neighborhood signatures, builds per-local-split probability dictionaries over masked label-conditioned signatures, and samples new snapshots during generation. The authors evaluate on seven SocioPatterns face-to-face datasets, comparing modularity, label assortativity, interaction duration, and PCA of neighborhood signatures against ETN-gen, and also propose and test CLETN and DLETN variants for label-absent settings. The central claim is that LETN reproduces community interaction structure substantially better than ETN-gen.","tokens_in":16597,"tokens_out":5879,"duration_ms":57594,"significance":"Face-to-face temporal network surrogates with community structure would be useful for privacy-preserving simulations of epidemics and social dynamics, and the proposed extension is conceptually simple and computationally light. The manuscript has strengths: seven public datasets, ten generations per dataset, explicit generation and validation rules, and a useful PCA diagnostic suggesting that LETN signatures separate communities. However, the validity of the central claim rests on metrics (modularity and label assortativity) that are direct aggregate summaries of the label-conditioned statistics the generator is fitted to reproduce; this circularity, together with the absence of hold-out tests, downstream simulations, or an attributed generative baseline, means that the paper does not currently establish realistic behavior beyond the training distribution. The work is a reasonable candidate after substantial validation improvements.","major_comments":[{"comment":"The main validation metrics are not independent of the fitted statistics. In Section 2.1, the probability dictionaries are computed from masked LETNS strings that include the ego node's label, the neighbor labels, and the masked last-layer bits, so the generator is explicitly fitted to label-pair transition frequencies; modularity computed with the original labels and label assortativity (Figures 6-7, Table 2) are direct aggregate summaries of those same label-pair frequencies. The strong match is therefore substantially built in, and the ETN baseline is expected to fail these metrics because it has no label channel. To support the claim of 'replicating realistic behaviors,' the paper should add at least one validation that does not reuse the training statistics, for example training on the first part of the observation period and evaluating on a held-out temporal segment, simulating a downstream process such as SIR epidemic spread and comparing outbreak statistics, or comparing against an attributed generator such as DyANE (reference [33]).","section":"Section 2.1 and Section 3, Figures 6-7, Table 2"},{"comment":"Table 3 reports Euclidean distances between snapshot-level metric curves for LETN, CLETN, and DLETN, but it does not include the ETN baseline, so it cannot support the claim that the labeled method improves temporal tracking; in addition, the snapshot-level values are temporally autocorrelated and the reported mean plus or minus error over ten runs is not a statistical test. Please add an ETN column and paired or bootstrap comparisons, or another appropriate test, to establish that the differences are meaningful.","section":"Section 3, Table 3"},{"comment":"The Hospital row of Table 2 shows that LETN does not improve over ETN on the headline metrics (original modularity 0.00 plus or minus 0.20 versus LETN -0.02 plus or minus 0.16 versus ETN -0.11 plus or minus 0.09; label assortativity -0.03 versus -0.07 versus -0.20). The blanket statement in the Introduction and Section 3 that the method 'significantly improves' on ETN-gen should therefore be qualified, and the paper should discuss when metadata labels do not align with community-structured interaction patterns.","section":"Section 3, Table 2 (Hospital row)"},{"comment":"For the CLETN and DLETN extensions, it is not stated whether the modularity and label assortativity in Table 3 are computed using the originally observed labels or the community-inferred labels. If the inferred labels are used for both the original and generated networks, the comparison is at risk of circularity because the community structure is estimated from the same aggregated data used to build the generator; a null model with shuffled labels or a comparison using the true metadata would clarify the added value.","section":"Section 2.3 and Table 3"}],"minor_comments":[{"comment":"The heatmaps are described as log-scaled but no colorbar is provided; please add a colorbar and specify whether the log scale is base 10 or natural logarithm.","section":"Figure 4, Panel B"},{"comment":"The mean interaction durations per community pair are shown for a single dataset without confidence intervals; adding error bars or a second dataset would make the duration-reproduction claim quantitative.","section":"Figure 5"},{"comment":"There is a typo in the table caption: 'T able 1' should read 'Table 1'.","section":"Page 7, Table 1"},{"comment":"The manuscript needs copyediting for grammar and phrasing, for example 'The advantages ... has led' in the abstract should be 'The advantages ... have led'.","section":"Abstract and Introduction"}],"recommendation":"major_revision","confidential_remarks":"The circularity concern is the main reason I cannot recommend acceptance; the contribution is useful but incremental. I would be willing to recommend favorably if the authors add non-circular validation (hold-out or downstream dynamics) and statistical comparisons, and if they qualify the Hospital result. The absence of a comparison with DyANE, which is cited in the related work and directly targets attributed dynamic networks, is a notable missed opportunity."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real news here is simple: the authors take ETN-gen and condition the egocentric temporal signatures on node labels, calling the result LETN. That is a modest but legitimate extension, and it works in the obvious sense. Across seven SocioPatterns datasets, LETN reproduces modularity and label assortativity far better than unlabeled ETN-gen, both at snapshot level and in the aggregated network. The PCA visualization and the interaction-duration heatmaps are genuinely informative and are more independent evidence than the headline metrics. The CLETN and DLETN variants, which infer labels when metadata are absent, are a useful practical addition, and the reported Euclidean distances in Table 3 suggest those variants track LETN closely.\n\nThe soft spot is the one the stress-test flags, and it is real. The probability dictionaries in Section 2.1 are built from label-conditioned transition frequencies, and the main validation metrics—modularity and label assortativity—are direct summaries of exactly those label-pair frequencies. So the central quantitative claim is partly circular: LETN is trained to match these statistics and then evaluated on them. The ETN baseline lacks labels and predictably fails the label-based metrics, so that comparison does not disconfirm the core mechanism. This does not make the method useless, and the paper is honest about its lineage, but the claim that surrogates 'replicate realistic behaviors' is stronger than the evidence supports. There is no hold-out temporal period, no downstream validation like epidemic or information diffusion, and no comparison with an attributed generator such as DyANE. The absence of code is a real hindrance for reproducibility.\n\nMy overall read: this is a solid engineering contribution, not a conceptual breakthrough. The extension is natural, transparently presented, and consistently demonstrated on public data. The circularity concern is significant enough that a referee should require an independent test—either a temporal hold-out, a downstream dynamics comparison, or a baseline that also uses labels—before accepting the full realism claim. But the method itself is sensible and likely useful for practitioners who want community-structured surrogate contact networks.\n\nWho gets value: researchers working on surrogate temporal networks, epidemic modeling, or privacy-preserving sharing of face-to-face interaction data. It deserves a serious referee; I would send it to review with a request for the additional validation rather than desk-reject it. I would cite it in my own work on temporal network surrogates, mainly as the label-conditioned extension of ETN-gen.","headline":"A clean, incremental extension of ETN-gen that adds node labels and reproduces community mixing statistics, but its headline validation is partly circular and needs an out-of-sample or downstream check.","tokens_in":17241,"tokens_out":661,"would_cite":true,"duration_ms":7989,"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":"Adding community labels to egocentric temporal patterns lets surrogate networks reproduce the community structure of real face-to-face interactions, matching modularity, label assortativity, and cross-community interaction durations…","keywords":["temporal networks","face-to-face interactions","surrogate networks","community structure","node labels","egocentric temporal neighborhoods","network generation","label assortativity"],"falsifier":"Generate LETN surrogates after training on only the first half of each dataset, then compute modularity, label assortativity, and mean inter-community interaction durations on the held-out second half; if the surrogate match degrades to the level of ETN-gen on that held-out period, the method has captured training-period patterns rather than transferable temporal behavior. Independently, simulate an epidemic on original and surrogate networks and compare outbreak size and timing; systematic divergence despite matching structural metrics would show the structural fidelity does not carry over to the dynamics the method is meant to support.","tokens_in":16182,"feed_emoji":"👥","tokens_out":5985,"duration_ms":55538,"temperature":0.7,"pith_summary":"Temporal interaction data from real settings are short, small, and privacy-sensitive, so researchers want surrogate networks that imitate them. This paper argues that previous surrogate generators, while reproducing local temporal patterns, ignore whether interacting nodes belong to different communities or roles, and therefore produce networks with near-random community structure. The authors extend a recent egocentric generator by labeling nodes (from metadata or from detected communities) and encoding those labels into each node's temporal neighborhood signature; the resulting Labeled ETN-gen (LETN) produces surrogates whose modularity, label assortativity, and cross-community interaction durations track the original networks across seven SocioPatterns datasets. If the claim is right, researchers can generate arbitrarily long, arbitrarily large synthetic contact networks with realistic within- and between-community mixing, without exposing real participants' data.","feed_headline":"Labels let surrogate social networks keep real communities","feed_subtitle":"Adding community labels to egocentric patterns matches modularity and interaction duration across 7 datasets.","key_machinery":"The load-bearing object is the Labeled Egocentric Temporal Neighborhood Signature (LETNS), a bit string that records, for each neighbor of an ego over a window of k snapshots, the neighbor's label in snapshots where a link exists and a 0 where it does not, with neighbor strings sorted lexicographically and the ego's own label prepended. These encoded signatures are masked over the final snapshot to form keys of probability dictionaries that store the empirical distribution of temporal extensions, with separate dictionaries per local split to capture daily periodicity. During generation, the algorithm samples desired links from these dictionaries for each ego, then validates the provisional layer by accepting bidirectional requests, randomly keeping half of unidirectional ones, and matching remaining stubs to maximize consensus. The label encoding is what carries community information: it lets the same local temporal pattern produce different link choices depending on who is interacting with whom, which is exactly what the unlabeled baseline cannot do.","core_discovery":"The central claim is that community and role labels, injected into the egocentric temporal neighborhood encoding, are sufficient to make surrogate temporal networks reproduce the community organization of face-to-face interaction data. Concretely, LETN computes, for each node and window of snapshots, a binary signature that includes the ego's label and the labels of neighbors across the window, sorts these neighbor signatures, and uses the masked signatures to build label-conditioned probability dictionaries per local time split. Sampling from these dictionaries during generation yields networks whose modularity and label assortativity distributions match the original across a primary school, three high schools, a hospital, and two workplaces, while the unlabeled ETN-gen baseline stays near zero modularity on temporal snapshots. The method also reproduces the average duration of interactions between specific community pairs, and, in the hospital, role-specific asymmetries such as nurses' many short contacts versus doctors' fewer long ones.","pith_inferences":["The paper's main validation metrics, modularity and label assortativity, are direct aggregations of the label-conditioned probabilities the generator encodes, so the match is partly circular; a stronger test would hold out the final days of each dataset and check whether the generated networks still match on that held-out period.","The authors do not report downstream dynamics; a natural extension is to run a simple SIR contagion on original and LETN-generated networks and compare epidemic sizes, peak times, and inter-community transmission rates.","Because LETN explicitly models label-conditioned mixing, it could be adapted to generate counterfactual contact networks (for example, rewiring cross-class contacts) to study what community configurations would do to spreading or opinion dynamics.","The CLETN and DLETN label-inference variants are validated only by snapshot-level Euclidean distances; an independent check would be to compare predictive accuracy on next-snapshot link prediction between LETN and ETN-gen."],"forward_implications":["Surrogate networks with realistic community structure can be generated at arbitrary temporal length and population size, so epidemic and opinion-dynamics simulations can run beyond the observed data collection window.","The generator works both with metadata-given labels and with labels inferred from community detection, so it applies to datasets without role or class annotations.","Because community structure emerges from local label-conditioned rules rather than a global layout constraint, the surrogates reproduce the time evolution of modularity, including the daily rise and fall in schools.","Role-specific mixing patterns, such as nurse-doctor versus nurse-patient contact durations, are preserved, enabling realistic hospital or workplace interaction simulations.","The method gives a direct way to test how community structure affects spreading processes by generating surrogates with matched egocentric statistics but different label mixing."],"supporting_citations":[{"why":"Defines the ETN-gen method that LETN extends; supplies the egocentric temporal neighborhood machinery and generation pipeline.","marker":"[38]"},{"why":"Provides the efficient procedure for mining egocentric temporal motifs that motivates the signature encoding used here.","marker":"[44]"},{"why":"Louvain community detection, used by the CLETN variant to assign labels to unlabeled datasets.","marker":"[45]"},{"why":"Primary school face-to-face interaction dataset with class labels, one of the validation datasets.","marker":"[47]"},{"why":"Hospital interaction dataset with role metadata (patients, nurses, doctors), used to demonstrate role-specific behavior reproduction.","marker":"[43]"},{"why":"High school interaction dataset with class labels, used to test the method on a setting with many classes.","marker":"[49]"},{"why":"Office workplace interaction dataset with role metadata, one of the validation datasets.","marker":"[51]"}],"fun_headline_variants":["Community labels let synthetic networks mirror real interactions","Labeled surrogates preserve modularity and interaction durations","Community-aware generation prevents loss of network structure","Adding node labels keeps surrogate temporal networks realistic"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that matching the label-conditioned statistics used to build the generator (modularity and label assortativity) counts as evidence of realistic network behavior; since those metrics are direct summaries of the same label interaction preferences the generator encodes, the validation may be circular and may not extend to unmeasured aspects such as dynamics on the network.","fun_headline_variants_meta":{"raw":{"variants":["Community labels let synthetic networks mirror real interactions","Labeled surrogates preserve modularity and interaction durations","Community-aware generation prevents loss of network structure","Adding node labels keeps surrogate temporal networks realistic"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000434,"raw_usage":{"total_tokens":2195,"prompt_tokens":914,"completion_tokens":1281,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":530,"completion_tokens_details":{"reasoning_tokens":1223}},"tokens_in":530,"tokens_out":1281,"duration_ms":10377,"temperature":1.0,"reasoning_tokens":1223,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:44:43.242763+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate LETN surrogates after training on only the first half of each dataset, then compute modularity, label assortativity, and mean inter-community interaction durations on the held-out second half; if the surrogate match degrades to the level of ETN-gen on that held-out period, the method has captured training-period patterns rather than transferable temporal behavior. Independently, simulate an epidemic on original and surrogate networks and compare outbreak size and timing; systematic divergence despite matching structural metrics would show the structural fidelity does not carry over to the dynamics the method is meant to support.","supporting_citations":[{"cited_title":"Data Mining and Knowledge Discovery 36(1), 355– 378 (2022) https://doi.org/10.1007/s10618-021-00803-2 18","cited_arxiv_id":null,"evidence_quote":"Provides the efficient procedure for mining egocentric temporal motifs that motivates the signature encoding used here."},{"cited_title":"PloS one 8(9), 73970 (2013) https://doi.org/10.1371/journal.pone.0073970","cited_arxiv_id":null,"evidence_quote":"Hospital interaction dataset with role metadata (patients, nurses, doctors), used to demonstrate role-specific behavior reproduction."},{"cited_title":"Network Science3(3), 326–347 (2015) https: //doi.org/10.1017/nws.2015.10","cited_arxiv_id":null,"evidence_quote":"Office workplace interaction dataset with role metadata, one of the validation datasets."}],"review_version":1}