{"id":"b709a452-c9bf-4e5b-a5fb-49e4aa4c136a","arxiv_id":"2607.15855","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"Temporal network generator that explicitly tunes community similarity and node turnover via UAMI-guided genetic search plus SBM edge generation.","lead":"A new generator builds temporal networks whose community structure (splits, merges, node churn) is controlled by a genetic algorithm guided by mutual information, then adds timestamped edges via a stochastic block model. A generalist might care because it gives researchers a tunable benchmark for testing dynamic community-detection algorithms under realistic churn.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"GA control loop is load-bearing and not established: the node-count mutation probabilities in §3.2.2 are inverted, and the only convergence evidence is two small trials; without a corrected/demonstrated search the 'explicit control' claim lacks support.","rationale":"The reader's weakest assumption (GA convergence) is the right general area, but I found a more specific, internal inconsistency in §3.2.2 that makes the described algorithm unable to steer node count in the intended direction. This reinforces rather than overturns the CONDITIONAL verdict: the authors need to fix the mutation probabilities and demonstrate convergence across a realistic parameter range. The connectivity lower bound and Markovian limitation are honestly discussed and are not the main risk. I would not reject outright because the error may be a typo and the fitness term could still let selection pull the population toward the target, but as written the method is not reproducible and the central control claim is unsupported.","tokens_in":13413,"tokens_out":9052,"duration_ms":100528,"concrete_test":"Reproduce the §3.2.2 mutation rules in isolation: start with n=400, nhat=500, and apply the stated premove/padd probabilities to a population; if the mean node count decreases instead of increasing, the sign error is confirmed. Then fix the signs and run a grid of at least 9 settings (n_start ∈ {200, 1000, 5000}; n_target/n_start ∈ {0.5, 1, 2}; s_target ∈ {0.2, 0.5, 0.8}) with 10 independent runs each, reporting the distribution of achieved UAMI and node count. If any setting misses the target by more than a tolerance (e.g., ±0.05 UAMI, ±5% nodes), the explicit-control claim fails as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The claim that users can dial a target inter-snapshot similarity and node count rests entirely on the genetic algorithm in §3.2. As written, the adaptive mutation probabilities are backwards: when current n < target nhat, the text removes assigned nodes (premove = max(0,(nhat-n)/nhat) > 0) and adds none (padd = max(0,(n-nhat)/nhat)=0), pushing the partition further from the target; the inequalities are reversed for n > nhat. Thus the described algorithm cannot perform the node-churn part of the advertised control. The only evidence that the GA reaches (s,n) targets is Figure 9, which shows two settings (n=400->500 and 500->400, s=0.5), with no code, no error bars, and no failure or scalability data. There is also no check that the converged partitions realize the intended split/merge events rather than achieving the UAMI score through label shuffling or node churn. Because every temporal edge depends on these partitions, an unreliable search undermines the central 'explicit control' claim, independent of the connectivity bound in §3.3.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a two-stage generative model for temporal networks with dynamic community structure. Stage one uses a genetic algorithm to build a sequence of node partitions in which (i) adjacent partitions have a prescribed UAMI similarity (a mutual-information measure for variable node sets) and (ii) each partition has a prescribed number of nodes; split, merge, add, and remove operators are used as mutations. Stage two generates timestamped edges via a stochastic block model with intra- and inter-community probabilities p_in and p_out, where p_out may be set from a lower bound that aims to keep the network connected. The model is applied to emails, citations, subreddit interactions, and NFT transactions; it is also used as a benchmark to study how node churn and community-change rate affect dynamic community detection algorithms. The central claim is that a user can explicitly dial the inter-snapshot community similarity and the node count while preserving connectivity.","tokens_in":13758,"tokens_out":8457,"duration_ms":74660,"significance":"The idea of controlling temporal community similarity directly through a similarity measure that handles changing node sets is attractive, and the two-stage design cleanly separates community evolution from edge/timestamp generation. If the genetic algorithm could be shown to reliably meet its targets, the model would fill a real gap: existing multilayer generators do not guarantee a specified similarity, and dynamic SBM extensions do not explicitly model node churn. The benchmark on churn and similarity is a useful contribution. The p_out lower-bound calculation is a reasonable start for connectivity control, though the claim as stated is stronger than what the mathematics proves. The paper would benefit from a fuller, less circular validation and a more rigorous convergence study; the present evidence does not yet support the abstract's 'explicit control' assertion. The underlying architecture is sound enough to be repaired.","major_comments":[{"comment":"The adaptive mutation probabilities are inverted. When n(t)_{i,j} < nhat(t), the text sets premove=(nhat-n)/nhat>0 and padd=0, so only removals are applied, driving the partition away from the target; when n>nhat, only additions occur. This contradicts the stated purpose and the convergence results in Figure 9, which cannot be produced by the algorithm as written. Please correct the formulas and confirm that the experiments were run with the corrected rule.","section":"§3.2.2"},{"comment":"The text says the bound 'guarantees global connectivity with high probability', but the union-bound argument only bounds the probability that some pair of communities has zero inter-community edges. It does not address isolated nodes within a community, and an edge between every community pair does not make the graph connected. Please revise the claim: at most this is a sufficient condition for no completely disconnected community pairs. If global connectivity is needed, add a condition on p_in or a per-node degree check.","section":"§3.3"},{"comment":"The real-data validation is partly self-referential. The UAMI similarities and node counts extracted from real networks are used as direct inputs to the generator, and the synthetic network is then evaluated by running Louvain and recomputing UAMI on the detected partitions. If the planted communities are detectable, the comparison is largely by construction. Please add a baseline control (e.g., randomizing the target similarities) and report the recovery accuracy of the planted partitions; preferably also compare statistics other than the fitted quantities.","section":"§4.2"},{"comment":"The only evidence for the central 'explicit control' claim is two single-run convergence experiments (Figure 9) for one similarity value (0.5) and two node-count transitions. There are no repeated runs, error bars, scalability tests, or failure cases. Moreover, the experiments do not verify that the final partitions realize the intended split/merge events rather than reaching the UAMI target through label shuffling or node churn. Please provide a systematic GA evaluation across the parameter ranges used in §4.3 and discuss failure modes.","section":"§4.4"},{"comment":"The definition of μ is inconsistent with the following sentence. μ=M_inter/(M_intra+M_inter) implies M_intra/M_inter=(1−μ)/μ, but the text states the ratio is μ/(1−μ). This inversion would change the interpretation of the benchmark sweeps. Please correct the formula and verify that the implementation matches the intended definition.","section":"§4.3"}],"minor_comments":[{"comment":"Figure 9 caption: 'shrank' should be 'shrunk'.","section":"§4.4"},{"comment":"In the D-DCSBM paragraph, 'yet it the model treats' is a typo; also 'absence nodes' should be 'absent nodes'.","section":"§2"},{"comment":"The line 'will makes the network broken' is ungrammatical. Also, the 'Taylor expansion' label is misleading: the derivation uses the expansion of −ln(1−p), not a Taylor approximation of the probability itself.","section":"§3.3"},{"comment":"The threshold for ignoring small communities is derived as if the two communities have similar size; clarify that the condition is on the product n_i n_j, and state how 'ignoring' is implemented.","section":"§3.3"},{"comment":"'empty model' should be 'null model'. Figure 3 contains duplicated text for 'Pair-to-pair similarity in community structures in the real temporal network'.","section":"§4.2"},{"comment":"The abstract's 'explicit control of community evolution including splits and merges' is stronger than what the model actually offers: the user can prescribe the UAMI similarity and node counts, but not the sequence or type of split/merge events. Please temper the wording.","section":"§3.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to build directly on the authors' own UAMI paper [39]. The GA section contains at least one internal inconsistency (inverted mutation formulas) and the μ definition is inverted relative to its own formula, so the experiments may have been run with a different implementation than described; the authors should be asked to reconcile the text with the actual code. No code or data availability statement is provided, which is problematic for a GA-based benchmark. The paper is within the scope of cs.SI and could become acceptable after a major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a decent generative framework, but the central 'explicit control' claim is weakened by a concrete error in the mutation probabilities and by thin convergence evidence. I'd send it to review, but the authors need to fix the search and sharpen the validation.\n\nWhat's new: the combination of UAMI from their prior work with a genetic search over partitions, explicit node churn, and timestamped SBM edge generation is a reasonable new tool. Unlike Bazzi et al. and dIRM, it gives a direct knob for inter-snapshot similarity and node count, and the benchmark on how churn hurts dynamic community detection is a genuinely useful result. The Discussion is honest about the Markovian limitation, which is a good sign.\n\nThe soft spots are real. Section 3.2.2 has the removal/addition probabilities inverted: when the current partition is smaller than the target, the text removes nodes, and when larger, it adds them. As written, the node-count control pushes away from the target rather than toward it. The two convergence plots in Figure 9 are the only evidence the GA works, and they cover only small cases with no code or error bars. That matters because every downstream edge inherits the partitions, so if the GA stalls, the whole 'dial a similarity' claim collapses. The connectivity bound in Section 3.3 is also overclaimed: it bounds the chance that some community pair has no inter-edges, not global connectivity of all nodes. Isolated nodes within a community are untouched.\n\nThe real-data validation is partly circular: node counts and UAMI similarities are extracted from Louvain on real networks, fed into the GA, and then compared on the same UAMI measure. There are no baselines or error bars, so 'reproduces' is partly by construction. And the split/merge control is never directly verified – the GA could hit the UAMI target through label shuffling or node churn instead of the intended events.\n\nThat said, the core design is coherent and the benchmark study is useful. The problems are fixable with a corrected GA, a few convergence tests at larger scale, and a validation that compares against a null model. I would assign a serious referee and ask for those changes.","headline":"A genuinely useful dynamic community generator with a real gap to fill, but the GA control loop has a concrete bug and the real-data validation is partly self-feeding; still deserves a referee.","tokens_in":14231,"tokens_out":2038,"would_cite":false,"duration_ms":19376,"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":"A generative model that turns a prescribed community-similarity trajectory into a timestamped temporal network with controlled node churn.","keywords":["temporal networks","generative model","dynamic communities","genetic algorithm","mutual information","node churn","stochastic block model","community detection benchmark"],"falsifier":"Run the genetic algorithm on a 1000-node source partition with target UAMI 0.7 and a prescribed node-count change of 20%; if, after a fixed iteration budget, the best candidate's UAMI remains below 0.6 or the node-count error exceeds 5%, the explicit-control claim for that regime is falsified.","tokens_in":13309,"feed_emoji":"🕸️","tokens_out":4253,"duration_ms":35726,"temperature":0.7,"pith_summary":"This paper proposes a two-stage generative model for temporal networks in which the experimenter can specify, in advance, how much the community structure changes between consecutive time steps and how many nodes are present at each step. The first stage runs a genetic algorithm that searches for a sequence of node partitions whose adjacent-step similarity, measured by a mutual-information variant that tolerates changing node sets, matches a prescribed target. The second stage converts those partitions into timestamped edges using a stochastic block model, with an inter-community connection probability bounded from below to keep the network connected. If the method works as claimed, it gives researchers a tunable null model and benchmark for testing dynamic community detection algorithms under controlled community-split rates and node churn.","feed_headline":"Dial community evolution speed in synthetic temporal networks","feed_subtitle":"New generator turns a target similarity into timestamped edges, with connectivity guaranteed.","key_machinery":"The load-bearing object is the union-adjusted mutual information (UAMI), a similarity score between two partitions that remains well-defined when the two node sets differ; it is used as the fitness target inside a genetic algorithm. The genetic algorithm's mutation operators are restricted to four community-evolution events — node addition, node removal, community splitting, and community merging — with add/remove probabilities adaptively adjusted toward the desired node count. A secondary but essential piece is the lower bound on the inter-community edge probability p_out, derived from Boole's inequality and a Taylor-expansion bound, which guarantees with probability at least 1-epsilon that","core_discovery":"The central claim is that community evolution in a temporal network can be generated to order: rather than letting similarity between snapshots emerge indirectly from hidden parameters, the model takes a target inter-snapshot similarity — the union-adjusted mutual information (UAMI) between consecutive partitions — and a target node count per snapshot as explicit inputs, then uses a genetic algorithm whose mutations are exactly the four elementary community events (node addition, node removal, community split, community merge) to find a partition sequence with the requested structure. Edges are then placed within each snapshot with intra-community and inter-community probabilities, where the","pith_inferences":["If the UAMI-guided genetic search scales, the same two-stage design could be applied to multilayer networks with millions of nodes by parallelising the population evaluation, a direction the paper mentions but does not implement.","The explicit similarity target opens a direct way to measure an algorithm's sensitivity to community-evolution speed: one could sweep target UAMI values at fixed churn and plot detection accuracy, effectively separating the effects of churn from the effects of merge/split rate.","Because the connectivity bound treats small communities as ignorable, the model embeds a decision about minimum community size that could be exposed as a parameter and tested for its effect on downstream detection."],"forward_implications":["Researchers can generate temporal networks with a prescribed rate of community change and node churn, which existing generators control only indirectly.","The same generator can be configured from real data (snapshot sizes and step-to-step similarities) to produce synthetic counterparts that reproduce short-range community-evolution statistics.","Dynamic community detection algorithms can be stress-tested against controlled churn: the paper shows increasing churn rate degrades detection accuracy, most strongly for random-walk-based methods.","The model offers a memoryless baseline: because only adjacent snapshots influence each other, deviation of a real network from this baseline reveals higher-order temporal dependencies."],"fun_headline_variants":["Mutual information steers synthetic community evolution","Generate temporal networks with controlled community dynamics","New model dials in community splits and merges","Synthetic networks with tunable community evolution","Community evolution on demand via mutual information"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole scheme rests on the genetic algorithm actually finding a partition that meets the requested similarity and node count; the paper's only evidence is two convergence runs (node additions and removals around 400–500 nodes), with no guarantee for larger or more complex settings.","fun_headline_variants_meta":{"raw":{"variants":["Mutual information steers synthetic community evolution","Generate temporal networks with controlled community dynamics","New model dials in community splits and merges","Synthetic networks with tunable community evolution","Community evolution on demand via mutual information"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000222,"raw_usage":{"total_tokens":1226,"prompt_tokens":616,"completion_tokens":610,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":360,"completion_tokens_details":{"reasoning_tokens":544}},"tokens_in":360,"tokens_out":610,"duration_ms":4746,"temperature":1.0,"reasoning_tokens":544,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T22:06:04.712612+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the genetic algorithm on a 1000-node source partition with target UAMI 0.7 and a prescribed node-count change of 20%; if, after a fixed iteration budget, the best candidate's UAMI remains below 0.6 or the node-count error exceeds 5%, the explicit-control claim for that regime is falsified.","supporting_citations":[],"review_version":1}