{"id":"149f66fe-569f-4cf6-aef7-16d774ed5119","arxiv_id":"1908.06940","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"The authors prove that spectral clustering on aggregated event counts consistently recovers communities in a new independent-pair Hawkes block model, and that moment estimators for the Hawkes parameters are consistent.","lead":"This paper proposes CHIP, a generative model in which every pair of users in a network exchanges messages according to its own bursty clock, with clock settings shared inside user communities. The paper proves that spectral clustering on total message counts and moment-based estimators are statistically consistent, and shows the method scales to tens of thousands of nodes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma B.1 substitutes Gaussian entries for Hawkes counts without any error bound; the spectral-norm bound underpinning Theorem 1 is therefore unproven, so the headline community-detection consistency does not yet have a rigorous proof.","rationale":"I read the paper as claiming two central results: consistent community detection by spectral clustering on the aggregated count matrix N, and consistent moment estimators for m=alpha/beta and mu. The estimation procedure and experiments are well executed, and the modeling contribution is valuable. However, the proof of the first result is not complete. Lemma B.1 replaces the true count matrix with a Gaussian matrix having the same entrywise variances, but the CLT for Hawkes counts is asymptotic in T and does not yield a valid non-asymptotic concentration inequality for the matrix spectral norm. In the sparse regime the paper explicitly targets (mu ~ log(n)/(nT)), counts are mostly Bernoulli-like and the Gaussian approximation is far from accurate, so the Bandeira-van Handel bound cannot be invoked as written. This is not a trivial gap: the entire misclustering bound in Theorem 1 depends on that spectral-norm control. A repair via Bernstein-type concentration for Hawkes counts is plausible, but it must be carried out and may change the rate or the required sparsity. The reader's weakest assumption identified both this Gaussian step and the finite-time correction to the mean. I agree that the Gaussian approximation is the more load-bearing issue; the finite-time correction is a finite-T bias that vanishes in the double limit and is secondary for the asymptotic theorems. The estimator theorem also overclaims consistency for beta (the paper itself says there are no guarantees), but this is an abstract-level overclaim rather than a flaw in the main spectral-clustering proof. Thus the paper needs a conditional acceptance with the proof of Lemma B.1 repaired; the reader's CONDITIONAL verdict is appropriate and I see no reason to move it.","tokens_in":27739,"tokens_out":15202,"duration_ms":162739,"concrete_test":"Re-derive Lemma B.1 without substituting Gaussians: use the Hawkes cluster representation (immigrant Poisson process with geometric offspring) to prove a non-asymptotic Bernstein-type bound on ||N - E[N]||_2 with the same sigma and sigma_* parameters as in (B.1)-(B.2). If the resulting bound has an extra log n factor or requires a stronger sparsity condition than the one used in Theorem 1, then the stated misclustering rate is not justified and Theorem 1 would need to be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof of Theorem 1 relies on Lemma B.1, which bounds ||N - E[N]||_2 by asserting that each count N_ij is asymptotically normal and then writing (N - E[N])_ij = g_ij sigma_ij with i.i.d. Gaussian g_ij (Supplementary, proof of Lemma B.1, after citing Hawkes & Oakes 1974). This is a substantive leap: the univariate CLT for each N_ij does not provide a joint Gaussian representation, and no Berry-Esseen-type rate or uniform error over the n^2 entries is supplied. In the sparse regime explicitly claimed to be covered (mu ~ log(n)/(nT), so E[N_ij] = O(log n / n)), the counts are far from normal (mostly 0/1), the Gaussian proxy is invalid, and the Bandeira-van Handel bound (Corollary 3.9) is not applicable to the actual count matrix without tail assumptions. Consequently, the misclustering bound in Theorem 1 is not rigorously established for the weighted adjacency matrix. The unweighted adjacency result (Theorem B.2) is not affected because it uses Lei-Rinaldo, which highlights that this is a specific gap in the central weighted-matrix argument.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces CHIP, a continuous-time network generative model in which each directed node pair follows an independent univariate exponential Hawkes process whose parameters depend only on the pair's communities. Estimation proceeds by spectral clustering on the aggregated count matrix N, followed by closed-form method-of-moments estimators for m = alpha/beta and mu, and a line search for beta. The main theoretical claims are that spectral clustering consistently recovers communities as n and T grow, with the misclustering rate bounded by k^2/(nT) in a simplified two-parameter setting, and that the estimators for m and mu are consistent and asymptotically normal. The paper also presents simulation experiments and comparisons on three real timestamped networks, emphasizing scalability to tens of thousands of nodes.","tokens_in":27972,"tokens_out":13869,"duration_ms":133751,"significance":"If the theoretical results were fully rigorous, this would be a notable contribution: it proposes a tractable Hawkes-based block model for timestamped relational data, gives the first consistency guarantees for community detection and parameter estimation in such a model, and provides a scalable implementation with reproducible code, simulation studies, and real-data comparisons against BHM and REM. The empirical scalability to the Facebook wall-post network is a clear strength. However, the core proofs have substantial gaps, so the guarantees as currently written are not established.","major_comments":[{"comment":"","section":"Supplementary B.1.3, proof of Lemma B.1 and Theorem B.1"},{"comment":"","section":"Section 4.2, Theorem 2 and Supplementary B.2.2"},{"comment":"","section":"Abstract, Section 3.2, Section 5.2"},{"comment":"","section":"Supplementary B.2.2, proof of Theorem 3"}],"minor_comments":[{"comment":"","section":"Supplementary B.1, Lemma B.1"},{"comment":"","section":"Supplementary B.1, finite-time correction"},{"comment":"","section":"Section 4.1, Theorem 1"},{"comment":"","section":"Section 5.2, Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The paper's theoretical core needs substantial revision before it is suitable for a journal. The Gaussian-substitution gap in Lemma B.1 and the incorrect variance in Theorem 2 are load-bearing and would likely be noticed by any careful reader. I recommend that the editor require a corrected proof of the spectral-norm bound, a corrected statement and proof of the asymptotic distribution in Theorem 2 (including the moment terms), and a careful restatement of the scope of the consistency claims. The empirical and algorithmic contributions are valuable, so I do not recommend outright rejection, but the current form does not meet the journal's standards for theoretical guarantees."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a useful paper with a genuine gap in the flagship proof. The CHIP model—independent Hawkes processes per node pair, sharing parameters within block pairs—is a natural and pragmatic variant of the Block Hawkes Model. The estimation pipeline (spectral clustering on the count matrix, then moment estimators for m and mu, line search for beta) clearly scales, and the real-data experiments are a genuine plus. The paper also delivers the first consistency-type analysis for a community point-process network model, which is a real contribution even if the proof isn't airtight.\n\nThe moment estimators for m and mu are standard delta-method stuff, and Theorem 2 is fine conditional on known or consistently estimated communities. The empirical MSE decay matches the predicted quadratic rate. So there's real substance here.\n\nNow the soft spots, in proportion. The big one is Lemma B.1. The proof simply asserts that N_ij is asymptotically normal and then writes (N - E[N])_ij = g_ij sigma_ij with i.i.d. Gaussian g_ij. That's not implied by a univariate CLT per entry, and no joint Gaussian approximation or Berry-Esseen-type error is supplied. In the sparse regime the paper explicitly claims (mu ~ log(n)/(nT)), the counts are mostly 0/1 and far from Gaussian, so the Bandeira–van Handel bound cannot be applied that way. That means Theorem 1, the headline community-detection consistency result, is not rigorously established as written. The unweighted adjacency result (Theorem B.2) is on firmer ground because it uses Lei–Rinaldo, which shows the gap is specific to the weighted case.\n\nTwo smaller issues. First, the abstract says \"consistent... estimators for the model parameters,\" but Section 5.2 admits the beta estimator has no guarantees. That's an overclaim; the paper should say m and mu, not all parameters. Second, the finite-time correction to the mean from Da Fonseca and Zaatour is dropped as negligible without quantifying the regime. Minor, but worth stating more carefully.\n\nThe central idea is plausible and the empirical work suggests the method works, but the proof of the central theorem currently rests on an unproven Gaussian substitution. That's fixable, and I hope the authors fix it.\n\nWho this is for: anyone working on point-process network models, especially with community structure. It deserves serious refereeing—the model, method, and experiments merit attention even if the proof needs revision. I'd send it out rather than desk-reject, but I'd demand a rigorous repair of Lemma B.1 and a toned-down abstract before publication.","headline":"Useful, scalable Hawkes block model with a real proof gap in the spectral clustering guarantee; worth serious refereeing but the headline consistency claim needs an honest fix.","tokens_in":28470,"tokens_out":1390,"would_cite":false,"duration_ms":17436,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60G55","62F12","62H30"],"pacs":[],"model":"deepseek-v4-flash","headline":"Spectral clustering on aggregated event counts is a consistent community detector as node count and observation time grow.","keywords":["CHIP model","Hawkes process","continuous-time network","community detection","spectral clustering","stochastic block model","moment estimator","consistency"],"falsifier":"Simulate the simplified CHIP model with known parameters at moderate $T$ and $\\mu \\asymp \\log(n)/(nT)$, then compare spectral-clustering adjusted Rand scores to the $k^2/(nT)$ rate and compare sample count means to $\\mu T/(1-\\alpha/\\beta)$. A systematic negative bias equal to the finite-time correction $-\\mu\\alpha(1-e^{-(\\beta-\\alpha)T})/(\\beta-\\alpha)^2$, or a misclustering rate that does not shrink as predicted, would show the asymptotic assumptions are not met in that regime.","tokens_in":27509,"feed_emoji":"🕸️","tokens_out":8296,"duration_ms":80245,"temperature":0.7,"pith_summary":"This paper introduces the Community Hawkes Independent Pairs (CHIP) model for networks whose edges are timestamped relational events, such as messages, emails, or financial transactions. The central claim is that a two-stage estimation recipe—aggregate events between each pair of nodes into a count matrix, run spectral clustering to recover communities, then estimate Hawkes process parameters from per-block-pair count means and variances—recovers both communities and model parameters consistently as the number of nodes and the observation time grow. The misclustering rate is bounded by a $k^2/(nT)$-type term, so longer observation windows can compensate for smaller networks. Because each node pair is modeled by an independent Hawkes process with shared block-pair parameters, the analysis is tractable and the estimator scales to networks with tens of thousands of nodes. The paper also reports that CHIP fits several real event datasets better than existing continuous-time network models.","feed_headline":"Spectral clustering provably finds communities in event networks","feed_subtitle":"A Hawkes model with independent node pairs gives consistent estimators that scale to Facebook-size networks.","key_machinery":"The load-bearing construction is independence between node pairs: for each block pair $(a,b)$, every ordered node pair $(i,j)$ with $c_i=a$, $c_j=b$ generates events as an independent univariate exponential Hawkes process with shared intensity parameters $\\mu_{ab}$, $\\alpha_{ab}$, $\\beta_{ab}$. This makes the aggregated count matrix $N$ a matrix of independent entries whose per-entry mean and variance are the Hawkes asymptotic moments $\\nu_{ab}=\\mu_{ab}T/(1-\\alpha_{ab}/\\beta_{ab})$ and $\\sigma_{ab}^2=\\mu_{ab}T/(1-\\alpha_{ab}/\\beta_{ab})^3$. The argument then uses three standard pieces: the ratio of sample mean to sample variance identifies $m=\\alpha/\\beta$ and $\\mu$; a sharp nonasymptotic bound on the spectral norm of a random matrix with independent entries controls $\\|N-\\mathbb{E}[N]\\|$; and a standard eigenvector-perturbation argument converts that control into a bound on the spectral-clustering misclustering rate. To recover $\\alpha$ and $\\beta$ separately, $\\hat{m}$ is plugged into the exponential-Hawkes log-likelihood, leaving a scalar line search over $\\beta$.","core_discovery":"The paper claims that for the CHIP model, spectral clustering applied to the weighted adjacency matrix $N$—the matrix of pairwise event counts—is a consistent estimator of community membership. In the simplified equal-size two-parameter version, the misclustering rate satisfies $r \\lesssim \\frac{k^2}{nT}\\frac{\\sigma_1^2}{(\\nu_1-\\nu_2)^2}$ as $n$ and $T$ grow, where $\\nu=\\mu T/(1-\\alpha/\\beta)$ and $\\sigma^2=\\mu T/(1-\\alpha/\\beta)^3$ are the asymptotic Hawkes count mean and variance. The paper further establishes that the moment estimators $\\hat{m}=1-\\sqrt{\\bar{N}/S^2}$ and $\\hat{\\mu}=\\frac{1}{T}\\sqrt{\\bar{N}^3/S^2}$ are consistent and asymptotically normal at rate $\\sqrt{n_{ab}}$, and gives an end-to-end mean-squared-error bound for estimating block-pair mean counts when communities are themselves estimated. In practice, the procedure is reported to fit a Facebook wall-post network with over 40,000 nodes and 800,000 events and to give better held-out-event log-likelihoods than prior models.","pith_inferences":["The moment estimators need only the count matrix, so they could serve as fast initialization or as inference when event timestamps are missing or too expensive to process—an application the paper does not itself propose.","Because the model assumes independent node pairs, reciprocity is captured only indirectly through symmetric block-pair parameters; adding an explicit reciprocity term would be a natural testable extension, which the paper's limitations discussion hints at.","The finite-time correction to the count mean quoted in the supplement implies a simple diagnostic: at moderate $T$, if sample means systematically fall below $\\mu T/(1-\\alpha/\\beta)$, the asymptotic regime behind the consistency theorems has not been reached.","The supplement's comparison of weighted versus unweighted adjacency matrices suggests a hybrid clustering rule—use the unweighted matrix when only $\\mu$ separates communities and the weighted matrix when burstiness carries the signal—that could outperform either alone in sparse networks."],"forward_implications":["As $T$ grows with $n$ and $k$ fixed, the misclustering rate bound falls like $1/T$, so observing the same network for longer yields the same accuracy gain as adding more nodes.","In sparse settings with $\\mu \\asymp \\log(n)/(nT)$, consistent recovery holds when the number of communities satisfies $k=o(\\sqrt{\\log n}\\,|c_1-c_2|)$.","The count matrix alone—without timestamps—suffices to estimate $\\mu$ and the branching ratio $m=\\alpha/\\beta$ with $\\sqrt{n_{ab}}$-consistent, asymptotically normal estimators, enabling closed-form confidence intervals.","With communities estimated by spectral clustering, the average mean-squared error for estimating block-pair mean counts decays at least linearly in $n/k$, not quadratically as it would with known communities.","Fitting the model to a 43,953-node, 852,833-event Facebook wall-post network takes about 141 seconds for $k=10$ blocks and yields better held-out log-likelihoods than the compared models."],"supporting_citations":[{"why":"Defines the stochastic block model whose block structure CHIP generalizes and which gives the rank-$k$ form of $\\mathbb{E}[N]$.","marker":"[15]"},{"why":"Introduces the Hawkes self-exciting point process used to generate event times in the CHIP model.","marker":"[16]"},{"why":"Provides the eigenvector-perturbation and misclustering bound for spectral clustering in stochastic block models used in the proof of the community-detection theorem.","marker":"[28]"},{"why":"Supplies the spectral-clustering consistency result and random-matrix control invoked for the unweighted adjacency matrix analysis.","marker":"[31]"},{"why":"Gives the asymptotic Hawkes count mean $\\nu=\\mu T/(1-\\alpha/\\beta)$ and variance $\\sigma^2=\\mu T/(1-\\alpha/\\beta)^3$ on which both the moment estimators and the spectral bound rest.","marker":"[56–58]"},{"why":"Supplies the sharp nonasymptotic bound on the spectral norm of a random matrix with independent entries used to control $\\|N-\\mathbb{E}[N]\\|$.","marker":"[62]"},{"why":"Introduces the Block Hawkes Model that CHIP simplifies by making node pairs independent while keeping the same number of parameters.","marker":"[9]"}],"fun_headline_variants":["CHIP model gives consistent community detection for event networks","Spectral clustering on event counts provably finds communities","Hawkes model with consistent estimators scales to big networks","Fits Facebook-scale networks with consistent community detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that, once $T$ is large, each pairwise event count is approximately Gaussian with mean $\\mu T/(1-\\alpha/\\beta)$ and variance $\\mu T/(1-\\alpha/\\beta)^3$; if observation time is too short or counts are too sparse, the dropped finite-time corrections become non-negligible and both the community-detection bound and the estimator consistency can fail.","fun_headline_variants_meta":{"raw":{"variants":["CHIP model gives consistent community detection for event networks","Spectral clustering on event counts provably finds communities","Hawkes model with consistent estimators scales to big networks","Fits Facebook-scale networks with consistent community detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001218,"raw_usage":{"total_tokens":4995,"prompt_tokens":914,"completion_tokens":4081,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":530,"completion_tokens_details":{"reasoning_tokens":4028}},"tokens_in":530,"tokens_out":4081,"duration_ms":26967,"temperature":1.0,"reasoning_tokens":4028,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:30:43.402292+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the simplified CHIP model with known parameters at moderate $T$ and $\\mu \\asymp \\log(n)/(nT)$, then compare spectral-clustering adjusted Rand scores to the $k^2/(nT)$ rate and compare sample count means to $\\mu T/(1-\\alpha/\\beta)$. A systematic negative bias equal to the finite-time correction $-\\mu\\alpha(1-e^{-(\\beta-\\alpha)T})/(\\beta-\\alpha)^2$, or a misclustering rate that does not shrink as predicted, would show the asymptotic assumptions are not met in that regime.","supporting_citations":[{"cited_title":"Consistent estimation of dynamic and multi-layer block models","cited_arxiv_id":null,"evidence_quote":"Defines the stochastic block model whose block structure CHIP generalizes and which gives the rank-$k$ form of $\\mathbb{E}[N]$."},{"cited_title":"Detectability thresholds and optimal algorithms for community structure in dynamic networks","cited_arxiv_id":null,"evidence_quote":"Introduces the Hawkes self-exciting point process used to generate event times in the CHIP model."},{"cited_title":"A survey of statistical network models","cited_arxiv_id":null,"evidence_quote":"Provides the eigenvector-perturbation and misclustering bound for spectral clustering in stochastic block models used in the proof of the community-detection theorem."},{"cited_title":"Pseudo-likelihood methods for community detection in large sparse networks","cited_arxiv_id":null,"evidence_quote":"Supplies the spectral-clustering consistency result and random-matrix control invoked for the unweighted adjacency matrix analysis."},{"cited_title":"The enron corpus: A new dataset for email classification research","cited_arxiv_id":null,"evidence_quote":"Supplies the sharp nonasymptotic bound on the spectral norm of a random matrix with independent entries used to control $\\|N-\\mathbb{E}[N]\\|$."},{"cited_title":"Junuthula, Maysam Haghdan, Kevin S","cited_arxiv_id":null,"evidence_quote":"Introduces the Block Hawkes Model that CHIP simplifies by making node pairs independent while keeping the same number of parameters."}],"review_version":1}