{"id":"760fd58a-ba84-4e2e-8a66-325c1fc50b57","arxiv_id":"2505.24534","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"HLSAD uses the singular values of Hodge Laplacians on temporal simplicial complexes to flag anomalous snapshots, outperforming graph-only baselines on several benchmarks.","lead":"This paper introduces HLSAD, a method that detects unusual events in networks that change over time by looking at the spectral signatures of higher-order interactions (triangles and beyond), not just pairwise links. It shows on synthetic and real datasets that this higher-order view catches anomalies that standard graph-based methods miss.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"HLSAD's sensitivity to localized anomalies is unvalidated: all synthetic anomalies are global parameter shifts, and top-ℓ Hodge spectra plus normalization can miss small perturbations.","rationale":"The reader's weakest_assumption identifies the same spectral-sensitivity issue: top singular values may not capture small structural changes orthogonal to the dominant spectral directions. I agree and sharpen it with two concrete mechanisms: the ℓ2 normalization of the concatenated singular-value vector, and the lack of any perturbation experiment in the paper. The paper's motivation (Section 5.1) cites Betti numbers and truncated-SVD optimality, but neither provides a stability guarantee for anomaly scoring; Weyl's inequality only gives an absolute bound on singular-value shifts, which becomes relatively negligible once normalized against large baseline spectra. The synthetic experiments all change generative parameters globally, so the detected anomalies are collective shifts in hundreds of simplices. The real-world ground-truth events (term boundaries, congressional polarization, social events) are likewise large-scale. Consequently, the evidence does not support a general claim of detecting arbitrary anomalies in temporal simplicial complexes. This is not a reason to reject: the method is a clean extension of LAD, the triangle-closing experiment does demonstrate a higher-order-only anomaly that graph Laplacians miss, the code is publicly available, and LAD is realized as a special case (K=0), providing a fair ablation. The appropriate action is to condition acceptance on either adding a localized-anomaly experiment or explicitly scoping the claim to global/collective anomalies. Since the reader already issued a conditional verdict, my read does not change that verdict.","tokens_in":17169,"tokens_out":9411,"duration_ms":115969,"concrete_test":"Generate two baseline sequences, each of 60 snapshots, from the triangle-closing model and from SBM plus clique lifting using the paper's parameters (Tables 3 and 6) with α=0.1. In each run, inject exactly one localized anomaly at a random interior time step: (a) add a single 2-simplex; (b) add a single edge and its induced triangles. Keep all generative parameters unchanged. Run HLSAD with w_s=5, w_l=10, K=2, and ℓ=10 and ℓ=50 per Laplacian, using the rank-based anomaly criterion. Repeat over 100 random seeds and report the fraction of runs in which the injected step is ranked first (Hits@1) and the mean score percentile of the injected step relative to normal steps. If the hit rate is not clearly above chance and the percentile is not above 95, the sensitivity assumption fails and the central claim must be restricted to global anomalies.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim presupposes that the top-ℓ singular values of the Hodge Laplacians are a sufficient and sensitive fingerprint of anomalous structural change (Section 5.1). This is not established for localized anomalies. The feature vector is the concatenation of the ℓ largest singular values per L_k, then ℓ2-normalized, and scored by angular distance from sliding-window averages (Eqs. 3-7). For a localized perturbation such as adding or deleting a single edge or triangle, the perturbation to L_k is a small-rank matrix with spectral norm O(1); by Weyl's inequality, each retained singular value moves by at most that norm. In a complex with thousands of simplices, the dominant singular values are large, so after normalization the relative change in the feature vector is tiny and can be swamped by normal temporal variance in the window baseline. The synthetic evaluation (Section 6.1, Tables 3-6) only injects global parameter shifts in SBM probabilities or triangle-closing probability, altering hundreds or thousands of simplices simultaneously; the real-world ground-truth events are also large-scale collective changes. No experiment perturbs a single simplex or small cluster. Thus the paper's headline contribution—anomaly detection in temporal sequences of simplicial complexes—is supported only for collective or global anomalies, not for the localized regime that the method's framing implies. This is a coverage gap in the evidence, not an internal contradiction.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HLSAD, a method for anomaly detection in temporal sequences of simplicial complexes. For each snapshot, HLSAD computes Hodge Laplacians up to a user-specified rank K, extracts the ℓ largest singular values from the up- and down-Laplacian components, concatenates and ℓ2-normalizes them into a feature vector, and assigns an anomaly score via the angular distance between the current vector and characteristic vectors computed from short- and long-term sliding windows. The method is evaluated on four synthetic settings (hybrid, resampled, large, triangle-closing) and three real-world datasets (UCI messages, Senate co-sponsorship, MIT Reality Mining), with comparisons to LAD, EdgeMonitoring, Activity Vector, TENSORSPLAT, and LetoChange. The paper claims to present the first methodology for anomaly detection in temporal sequences of simplicial complexes and reports that the simplicial representation yields improved accuracy and efficiency over graph-only baselines.","tokens_in":17428,"tokens_out":8098,"duration_ms":87570,"significance":"If the claims are established, the paper fills a genuine gap by extending spectral graph anomaly-detection methods to higher-order simplicial structures, and it provides a public implementation. The experimental results suggest that Hodge-Laplacian spectra can be more discriminative than graph spectra in settings where group interactions carry the signal, and the method inherits desirable properties such as permutation invariance and tolerance to varying node sets. However, the evaluation has serious weaknesses: the Hits@N metric is used inconsistently, no replicate runs or error bars are reported, hyperparameters appear to be selected per dataset without a validation protocol, and the sensitivity to localized anomalies is never tested. These issues currently preclude accepting the paper's broad claims.","major_comments":[{"comment":"The Hits@N metric is defined as the proportion of the top-N detected points that are true anomalies (i.e., precision@N), and the text notes that when N equals the number of ground-truth anomalies, Hits@N equals recall. Under this definition, the reported Hits@10 = 1.0 for the UCI dataset, which has only two labeled events, is impossible (the top-10 list cannot consist entirely of true anomalies when only two exist). Likewise, the Figure 4 caption describes the dashed line as 'Hits@N+2 ... accuracy with up to 2 false positives,' which is a recall-style interpretation and cannot reach 1.0 under the precision definition. The paper must clarify whether Hits@N is precision@N, recall@N, or a binary 'all anomalies in top N' success indicator, and all reported values and captions must be made consistent with that choice. This is load-bearing because every quantitative comparison in the paper rests on this metric.","section":"Section 6, first paragraph; Figure 4 caption; Table 1"},{"comment":"No replicate runs, random seeds, or error bars are reported for any synthetic or real-world experiment. The datasets generated from the stochastic block model and the triangle-closing process are stochastic, and the real-world ground-truth event lists are fixed, yet all plots and tables appear to reflect a single run. In the resampled setting, for example, the paper states that LAD reaches Hits@7 = 0.8 with 250 singular values while HLSAD requires about 320 singular values; without multiple trials it is impossible to know whether this difference is meaningful. The authors should report means and standard deviations over multiple random seeds, or justify why a single run suffices.","section":"Section 6.1 and Figure 4"},{"comment":"The method has several free parameters (maximum rank K, number of singular values ℓ, short and long window sizes w_s and w_l, anomaly threshold τ or number of top detections k, and the choice of up versus down Laplacian components). The paper does not describe a validation protocol for selecting these parameters. Section 6.3 reports Senate results 'using only the top 6 singular values,' Section 6.2 fixes UCI windows 'following the arguments outlined in previous works,' and Appendix C shows that performance varies substantially with window size and states that 'the context window sizes can have significant influence on the performance of HLSAD' and that one should 'fine-tune the context window size specifically for the dataset at hand.' This suggests that parameters may have been chosen on the labeled test data, which would make the reported superiority circular. Please specify exactly how each parameter was selected, ideally through a training/validation split, and provide a sensitivity analysis over the parameter ranges.","section":"Sections 5.2, 6.2, 6.3, and Appendix C"},{"comment":"The evaluation probes only collective or global anomalies. All synthetic injections in Tables 3, 4, and 6 are SBM probability shifts or triangle-closing probability changes affecting hundreds or thousands of simplices simultaneously, and the real-world ground-truth events are large-scale collective changes. No experiment perturbs a single simplex or a small local cluster. The feature vector consists of the ℓ largest singular values of the Hodge Laplacians, which by Weyl's inequality are Lipschitz in the perturbation; a small-rank, small-norm perturbation may be invisible to the top-ℓ spectrum, especially after ℓ2-normalization and angular-distance scoring. The paper should either restrict its claims to collective events and change points or add a synthetic experiment with localized anomalies (e.g., adding or deleting a single edge, triangle, or small cluster) to demonstrate that the method actually detects the localized regime that the title and framing imply.","section":"Section 6.1 and Section 5.4"}],"minor_comments":[{"comment":"The definition of rank k_max as 'max{|S^k| | S^k ∈ X}' uses the cardinality of the largest simplex rather than its dimension; the equation should be k_max = max(|S^k| - 1).","section":"Section 3"},{"comment":"The baseline is spelled 'TENSORPLAT' in Table 1 but 'TENSORSPLAT' in the text; please use a consistent spelling.","section":"Table 1 and Section 2"},{"comment":"The word 'infered' should be 'inferred'.","section":"Appendix C"},{"comment":"In the row for step 92, the inter-community probability is listed as '0.00755', which appears to be a typo for '0.0075'.","section":"Table 4"},{"comment":"The complexity expression O(T * sum_k n_k^2 log ℓ) should define n_k explicitly as the number of k-simplices and should clarify whether sparsity is exploited; the preceding sentence gives an O(n^2 log ℓ) bound for dense matrices, which is not the same as a sparse SVD cost.","section":"Section 5.5, Eq. (8)"},{"comment":"The abstract and conclusion claim 'enhanced computational efficiency,' but no wall-clock timing comparison is reported; the efficiency argument relies only on the number of singular values needed, so please add a direct runtime comparison or soften the claim.","section":"Section 7"},{"comment":"The availability link already lists a public URL, while the text in Appendix B says the code 'will be made publicly available with the camera-ready version'; please align these statements.","section":"KDD Availability Link and Section 6"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely topic and the public code is a strength, but the evaluation needs substantial revision before the claims can be accepted. The Hits@N inconsistency is likely fixable by relabeling or recomputing numbers, but the lack of replicate runs and the absence of a validation protocol are more fundamental and will require new experiments. The localized-anomaly gap is a scope issue that should be made explicit or addressed empirically. I would not reject the paper, but it needs major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear Colleague,\n\nBottom line: this is a well-executed, honest extension of LAD to Hodge Laplacians and the first anomaly-detection method for temporal simplicial complexes. The central empirical claim—that higher-order spectra beat graph-only spectra when the signal lives in group interactions—holds up in the tests they ran. It deserves a proper peer review.\n\nWhat's new: the Hodge Laplacian feature construction (up/down components, rank-truncated singular values, zero-padding for low-rank complexes) and the triangle-closing synthetic benchmark that forces anomalies to appear in 2-simplices only. The method itself is a straightforward port of the Akoglu–Faloutsos/Huang et al. framework, but the simplicial domain is a genuine gap and they are the first to fill it. The code is public (Zenodo and RWTH GitLab), which makes the paper reproducible as stated.\n\nThe soft spots are real but not fatal. The Hits@N definition is internally inconsistent: it first reads as precision, then claims equivalence to recall. That needs to be clarified. Second, the sensitivity of top-ℓ Hodge spectra to localized perturbations is untested. All synthetic anomalies are global parameter shifts moving hundreds of simplices; the real-world events are large-scale too. A single-simplex flip would likely be lost after ℓ2 normalization, but no experiment probes that regime. Third, there are no error bars or replicate runs, and hyperparameters (ℓ, window size, K) appear to be tuned per dataset without a validation protocol. These are the things to push on in revision, not reasons to reject.\n\nThe reader's circularity concern is fair but mild: they don't fit a model to labeled anomalies, so the main burden is the hyperparameter selection, which is indeed under-described.\n\nWho this is for: anyone working on temporal higher-order networks, epidemic modeling, or spectral graph mining. It's not a breakthrough, but it's a useful, credible first tool.\n\nRecommendation: send to peer review. Fix the metric definition, add variance estimates, and at least discuss the localized-anomaly blind spot. I'd cite it if I worked in this area.\n\nBest","headline":"A clean first step for anomaly detection on temporal simplicial complexes; higher-order spectra help in the tested regimes, but the evaluation has a metric muddle and an unprobed localized-anomaly blind spot.","tokens_in":18000,"tokens_out":3253,"would_cite":true,"duration_ms":37242,"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":"HLSAD uses Hodge Laplacian spectra to detect anomalies in time-evolving simplicial complexes, outperforming graph-only baselines on higher-order and lifted data.","keywords":["anomaly detection","change point detection","event detection","simplicial complex","Hodge Laplacian","spectral methods","higher-order interactions","graph lifting"],"falsifier":"Construct a synthetic sequence in which the only anomaly is a small, localized rewiring — say, moving a single triangle from one community to another in a 500-node stochastic block model — so that the top few singular values of $\\mathbf{L}_0,\\mathbf{L}_1,\\mathbf{L}_2$ change by less than numerical noise. If HLSAD's score never exceeds a threshold calibrated on the normal segment for that snapshot, then the top-$\\ell$ singular-value fingerprint is not sufficient for the general sensitivity the paper claims.","tokens_in":16895,"feed_emoji":"🔺","tokens_out":7374,"duration_ms":87569,"temperature":0.7,"pith_summary":"This paper introduces HLSAD, a method for detecting anomalies in sequences of simplicial complexes — sets of points together with triangles and higher-order cells, not just edges. The authors aim to establish that the spectrum of Hodge Laplacians, computed for each snapshot and tracked through sliding windows, is a better fingerprint of structural change than the ordinary graph Laplacian spectrum. They argue that many real systems carry information in multi-way interactions, either natively or after lifting a graph to its clique complex, and that graph-only detectors are blind to changes that live in those higher-order cells. If the claim holds, HLSAD offers a practical way to monitor dynamic higher-order networks — epidemic contact groups, legislative co-sponsorship, proximity interactions — and to improve detection on ordinary dyadic data by enriching it with simplicial structure.","feed_headline":"Hodge Laplacian spectra catch anomalies graph-only methods miss","feed_subtitle":"Scoring snapshots by Hodge Laplacian spectra surfaces multi-way interaction changes that dyadic baselines miss.","key_machinery":"The load-bearing object is the spectrum of the combinatorial Hodge Laplacians $\\mathbf{L}_k$, whose kernel has dimension equal to the $k$-th Betti number, so the spectrum encodes topological features such as components, cycles, and higher holes. The key trick is to look at both components of each Laplacian: the down-Laplacian $\\mathbf{B}_k^\\top\\mathbf{B}_k$ records how $k$-simplices decompose into $(k-1)$-simplices, while the up-Laplacian $\\mathbf{B}_{k+1}\\mathbf{B}_{k+1}^\\top$ records how they combine into $(k+1)$-simplices. Singular values are invariant under reordering of vertices and simplices, so the feature vector works without node alignment, and truncated SVD on sparse Laplacians keeps the cost at $O(T \\sum_{k=0}^{K} n_k^2 \\log \\ell)$ with randomized SVD.","core_discovery":"Formally, HLSAD computes the combinatorial Hodge Laplacians $\\mathbf{L}_k = \\mathbf{B}_k^\\top\\mathbf{B}_k + \\mathbf{B}_{k+1}\\mathbf{B}_{k+1}^\\top$ for ranks $0$ through $K$ at each time step, extracts the $\\ell$ largest singular values from each (padding with zeros when the complex is thin), and concatenates them into a feature vector $\\boldsymbol{\\sigma}^{(t)}$. A context window of past snapshots is summarized by the leading left singular vector of the matrix of their normalized spectra, and the current snapshot is scored by its angular distance to the short-term and long-term typical spectra. The final anomaly score is the maximum of the two deviations, with isolated high scores classified as events and persistent elevations as change points. The central claim is that this higher-order spectral fingerprint detects anomalies that graph-only spectral methods miss, and that on clique-lifted dyadic data it reaches comparable accuracy with fewer singular values, hence lower cost.","pith_inferences":["A stress test the paper does not run is a localized anomaly that leaves the top-$\\ell$ singular values nearly unchanged; if that score stays below threshold, detection sensitivity is bounded by the spectral gap rather than by structural change.","Because the down-Laplacian at rank $k+1$ records how $(k+1)$-simplices attach to $k$-simplices, HLSAD could in principle flag the onset of higher-order community formation before it is visible in the graph skeleton; this is an untested extrapolation.","The concatenated up/down singular values form a general simplicial-temporal feature vector that could be fed to any downstream classifier or change-point model, turning HLSAD into a feature extractor rather than a standalone detector."],"forward_implications":["An anomaly that changes only triangle-level structure, with no edge-level change, still moves the score because the rank-2 Hodge Laplacian spectrum is part of the feature vector.","On dyadic data, lifting the graph to its clique complex before scoring yields higher Hits@N with fewer singular values, so the higher-order representation is more accurate and cheaper per unit of accuracy.","Because the score takes the maximum of short- and long-window deviations, isolated spikes are classified as events while sustained elevations are classified as change points, without fitting a second model.","The method needs no node alignment and no fixed node count, so it applies to growing networks and to datasets where simplex identities cannot be tracked across time.","On the tested real datasets, the method reaches perfect Hits@10 on the UCI messages data and Hits@2 on the Senate co-sponsorship data, and on the Reality Mining data it matches the model-based baseline's precision at zero delay while exceeding its recall."],"supporting_citations":[{"why":"Supplies the spectral-pattern idea and the SVD-based typical spectrum that HLSAD adopts for its context windows.","marker":"[1]"},{"why":"Defines the graph Laplacian baseline and the hybrid and resampled synthetic settings that HLSAD replicates and compares against.","marker":"[21]"},{"why":"Provides the continuity-rate model, the EdgeMonitoring comparison, and the Senate anomalies used as ground truth.","marker":"[33]"},{"why":"Contributes the model-based LetoChange baseline and the sixteen Reality Mining events used as ground truth.","marker":"[27]"},{"why":"Establishes that the Hodge Laplacian kernel dimension equals the Betti numbers, motivating spectral features as topological descriptors.","marker":"[9]"},{"why":"Gives the optimal low-rank approximation result that ties the ordered singular values to structural information loss.","marker":"[8]"},{"why":"Supplies the randomized SVD algorithm that gives HLSAD its low computational complexity per Laplacian.","marker":"[16]"},{"why":"Motivates clique lifting by showing that simplicial complexes can be strictly more discriminative than their graph skeletons.","marker":"[3]"},{"why":"Provides the UCI online message dataset and its known term-boundary events used in the real-world evaluation.","marker":"[26]"},{"why":"Provides the MIT Reality Mining proximity dataset whose Bluetooth contacts are lifted and scored against external events.","marker":"[7]"}],"fun_headline_variants":["Hodge Laplacians beat graphs at anomaly detection","Simplicial spectra catch higher-order change points","HLSAD: Hodge spectra spot multi-way anomalies","Graphs miss it: Hodge Laplacians see anomalies"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the top $\\ell$ singular values of the first $K$ Hodge Laplacians form a complete fingerprint, so that every meaningful structural anomaly changes at least one of these values enough to move the angular score.","fun_headline_variants_meta":{"raw":{"variants":["Hodge Laplacians beat graphs at anomaly detection","Simplicial spectra catch higher-order change points","HLSAD: Hodge spectra spot multi-way anomalies","Graphs miss it: Hodge Laplacians see anomalies"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000179,"raw_usage":{"total_tokens":1266,"prompt_tokens":877,"completion_tokens":389,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":493,"completion_tokens_details":{"reasoning_tokens":324}},"tokens_in":493,"tokens_out":389,"duration_ms":5277,"temperature":1.0,"reasoning_tokens":324,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:20:03.978820+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a synthetic sequence in which the only anomaly is a small, localized rewiring — say, moving a single triangle from one community to another in a 500-node stochastic block model — so that the top few singular values of $\\mathbf{L}_0,\\mathbf{L}_1,\\mathbf{L}_2$ change by less than numerical noise. If HLSAD's score never exceeds a threshold calibrated on the normal segment for that snapshot, then the top-$\\ell$ singular-value fingerprint is not sufficient for the general sensitivity the paper claims.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the spectral-pattern idea and the SVD-based typical spectrum that HLSAD adopts for its context windows."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the graph Laplacian baseline and the hybrid and resampled synthetic settings that HLSAD replicates and compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the continuity-rate model, the EdgeMonitoring comparison, and the Senate anomalies used as ground truth."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contributes the model-based LetoChange baseline and the sixteen Reality Mining events used as ground truth."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes that the Hodge Laplacian kernel dimension equals the Betti numbers, motivating spectral features as topological descriptors."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Motivates clique lifting by showing that simplicial complexes can be strictly more discriminative than their graph skeletons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the MIT Reality Mining proximity dataset whose Bluetooth contacts are lifted and scored against external events."}],"review_version":1}