{"id":"4775e253-cee7-4b8e-b0ef-187dae76954c","arxiv_id":"2501.04529","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A Bregman ADMM module imposes sparse and low-rank structure on responsibility and attention matrices in temporal point processes, improving performance and interpretability of event branch inference.","lead":"This paper introduces a plug-and-play module that makes event-triggering matrices in temporal point process models sparse and low-rank, yielding clearer event branches. The authors report improved prediction and interpretability on synthetic and real event-sequence data.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing ground-truth branch evaluation: synthetic Hawkes data has known branches, yet only likelihood/accuracy is reported, leaving the central interpretability claim unsupported.","rationale":"I read the paper's central claim as two-fold: (i) plugging BADMM improves predictive performance; (ii) it provides interpretable structured event branches. The first is supported by Table 2, though improvements vary across datasets. The second is the paper's distinctive contribution, but its evidence is qualitative. The synthetic dataset is ideal for ground-truth evaluation and is not used for that purpose. This is the most load-bearing gap because it directly targets the claim's novelty. The reader's identified EM convergence issue is real but secondary: it concerns the training procedure, whereas the branch-quality issue concerns the output itself. I partially agree with the reader; their concern would also be worth testing via checking monotonicity of observed likelihood, but I recommend adding ground-truth branch recovery as the primary check. The verdict remains CONDITIONAL because the deficiency is addressable with additional experiments, and the existing empirical results still provide some support for the method's practical utility.","tokens_in":14429,"tokens_out":4750,"duration_ms":48939,"concrete_test":"On the synthetic Conttime dataset, which is simulated by a Hawkes process, record the true parent/triggering assignments for each event during simulation. After training HP with BADMM (both R=||.||_* and R=||.||_{1,2}), compute the final structured responsibility matrix B for held-out sequences. Evaluate branch recovery by thresholding/sorting B and comparing to ground-truth parent indicators: report AUC, precision@k (top triggers per event), and normalized mutual information. Compare against the unregularized EM responsibility matrix and the Sinkhorn baseline. If BADMM does not significantly improve branch recovery over unregularized EM, the interpretability claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the BADMM module infers interpretable structured event branches. The only branch-level evaluation is the '12 Angry Men' juror ranking, where GPT-4o serves as pseudo-ground-truth, not an actual branching structure. Crucially, the synthetic Conttime dataset (Table 1) is generated by a Hawkes process, so the true event transition/branching structure is known. No experiment compares the inferred B (or attention map) against this ground truth. Without such a comparison, the sparse and low-rank visualizations in Figure 3 only demonstrate regularization, not correctness: the module may remove true triggering relations or fabricate false ones. The performance improvements in Table 2 could arise from the regularizer alone, independent of branch fidelity. The paper's own limitation section states the authors 'plan to verify the rationality of our method from the viewpoint of Statistics,' conceding that statistical validity is currently unaddressed. This gap is load-bearing because if branch recovery is no better than the unregularized responsibility matrix, the main novelty collapses to a generic sparsity regularizer. The EM substitution concern raised by the reader compounds this: replacing R(t) with a KL-regularized B(t) in the E-step without a bound-preserving argument means the structured matrix is not even guaranteed to be a posterior; but the deeper issue is that no evidence shows it is a better branch estimator. A quantitative branch-recovery test on Conttime would settle this.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a plug-and-play Bregman ADMM (BADMM) module for temporal point processes (TPPs) that imposes sparse and low-rank structure on responsibility matrices in the EM algorithm for classical Hawkes processes and on attention maps in Transformer-based neural TPPs. The resulting matrices are interpreted as event transition matrices encoding hidden branching structure. The module is implemented by unrolling Bregman ADMM iterations, with two regularizer choices: nuclear norm and sparse group-lasso. Experiments on one synthetic and five real-world datasets report log-likelihood and accuracy, along with qualitative visualizations of inferred transition matrices and a case study on the movie '12 Angry Men'.","tokens_in":14768,"tokens_out":8321,"duration_ms":74695,"significance":"The idea of a generic, optimization-driven module that can be plugged into both classical and neural TPPs to regularize inferred event transition matrices is relevant and timely. The paper provides code, which is a strength. If the branch-inference claim were quantitatively validated, the module could be a useful component for interpretable TPP modeling. However, the current evidence is largely qualitative, and the theoretical grounding for the EM variant is absent. The contribution is therefore more of an empirical engineering proposal than a fully validated method.","major_comments":[{"comment":"The paper's central claim is that the BADMM module infers interpretable structured event branches, but no experiment compares the inferred transition matrix (B or the attention map) against ground-truth branching structure. The Conttime dataset (Table 1) is simulated by a Hawkes process, so the true parent/triggering relations are known; however, Table 2 reports only ELL and ACC, and Figure 3 provides only visualizations. A quantitative branch-recovery evaluation (e.g., precision/recall of inferred parent assignments, AUC of predicted triggering probabilities, or similarity to the true transition matrix) is necessary to support the interpretability claim. Without it, the reported performance gains could result from the regularizer alone, independent of branch fidelity, and the core novelty of the paper is not empirically established.","section":"Experiments, 'Effectiveness and Rationality'"},{"comment":"The paper states that for a Hawkes process, the BADMM algorithm can be applied 'directly in the E-step' to obtain a structured responsibility matrix. In classical EM, the responsibility matrix R(t) in Eq. (4) is the posterior expectation of the latent branching variables, and its use guarantees that Q(θ, θ(t)) is a lower bound on the observed-data log-likelihood. Replacing R(t) with the minimizer of Eq. (7) breaks this lower-bound property, so monotonic likelihood increase and convergence to a stationary point of the observed-data likelihood are not guaranteed. No theoretical or empirical justification (e.g., a monotonicity check, a modified objective, or a convergence experiment) is provided. The limitation section explicitly states that statistical verification is left to future work. This is load-bearing for the Hawkes experiments and for the claim that the method operates within the maximum likelihood estimation framework.","section":"Proposed Method, 'BADMM Module for Structured Event Branches', final paragraph"},{"comment":"In the update for X2 when R(·) = ∥·∥1,2, the paper gives x_{n,2}^{(t+1)} = τ_n S_{λα/ρ}(b_n + z_{n,2}) with τ_n = (1 − (1−α)λ/(ρ∥S_{λα/ρ}(b_n + z_{n,2})∥₂))_+. This is the proximal operator for the combined penalty α∥·∥1 + (1−α)∥·∥1,2, but in the splitting of Eq. (8) the ℓ1 term is already assigned to X1; the X2 subproblem should use only (1−α)∥·∥1,2, which would yield τ_n = (1 − (1−α)λ/(ρ∥b_n + z_{n,2}∥₂))_+ without the inner soft-threshold. As written, the algorithm double-counts the ℓ1 regularizer, changing the objective being solved. Please correct the formula or clarify the intended splitting.","section":"Proposed Method, 'BADMM Module for Structured Event Branches', update equations"},{"comment":"The claim that the BADMM module 'consistently improves' performance is contradicted by several entries. For example, SAHP ACC on Taobao is 0.434 for Softmax, Sinkhorn, BADMM1,2, and BADMM*; THP ELL on Amazon is 0.542 for Softmax, BADMM1,2, and BADMM*; and many BADMM results are within one standard deviation of the baseline. The improvements are consistent only for the HP backbone on ELL, not across all settings. The paper should either report statistical significance tests (e.g., paired tests across the three seeds) or moderate the claim to reflect that gains are present in most but not all configurations.","section":"Experiments, Table 2"}],"minor_comments":[{"comment":"The definition of the nuclear norm uses 'σN (B)' which should be 'σ_n(B)' (sum of singular values).","section":"Proposed Method, Eq. (7)"},{"comment":"The URL 'https://github.com/qingmeiwangdaily/BADMM TPP' contains a space; the correct URL should be 'BADMM_TPP'.","section":"Abstract"},{"comment":"'Sinkhormer' is a typo for 'Sinkformer'.","section":"Related Work"},{"comment":"The subfigure labels and color bars are small; consider enlarging them for readability.","section":"Experiments, Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript presents a useful plug-and-play optimization module, but the missing branch-recovery evaluation and the unproven EM modification are significant gaps. The X2 update formula issue may indicate an implementation error. These concerns are addressable within the paper's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a genuine, reusable plug-and-play module—Bregman ADMM unrolled into a layer that imposes sparse/low-rank structure on responsibility matrices and attention maps for TPPs. The code is out, the updates are closed-form, and the experiments are honest about modest gains. What is missing is any quantitative test of the central branch-inference claim.\n\nThe new element is the specific combination: lower-triangular row-stochastic constraint plus Bregman ADMM with two choices of low-rank regularizer (nuclear norm and sparse group lasso), applied to both EM-based Hawkes and transformer TPPs. That integration is not in the cited literature. The paper does a good job positioning against Sinkhorn-based modules and explaining why those break the lower-triangular structure. They release code, so the work is reproducible.\n\nThe soft spots are real but addressable. The biggest one: the synthetic Conttime dataset is generated by a Hawkes process, so the ground-truth event branches are known. The paper only reports likelihood and accuracy; there is no comparison of inferred B against the true transition matrix. Without that, the interpretability claim rests on visualizations and the 12 Angry Men ranking, where GPT-4o serves as pseudo-ground-truth. That is not branch-truth. If the module recovers branches no better than the unregularized responsibility matrix, the novelty collapses to a generic regularizer. A simple precision/recall or rank-correlation experiment on Conttime would settle this.\n\nAlso, for the Hawkes EM variant, replacing the posterior responsibility R(t) with the KL-regularized B(t) in the E-step is not justified by monotonicity or convergence. This is fine if treated as a regularized heuristic, but the paper should say so and empirically verify that it doesn't break EM. The performance gains are often small and sometimes flat (e.g., SAHP ACC on Taobao is 0.434 for all variants), so the contribution is more about interpretability than accuracy.\n\nWho this is for: anyone working on TPP interpretability or optimization-based neural layers. It deserves a serious referee: the idea is sound, the code is available, and the gap is a missing experiment rather than a fatal flaw. I'd recommend a conditional accept with a required branch-recovery evaluation and a clear statement about the EM substitution.","headline":"A practical BADMM module that regularizes TPP responsibility/attention matrices into sparse low-rank branches, with solid engineering but no quantitative evidence that the inferred branches match ground truth.","tokens_in":15282,"tokens_out":2978,"would_cite":true,"duration_ms":28520,"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 Bregman-ADMM module that imposes sparse and low-rank structure on event-transition matrices improves both the accuracy and the interpretability of temporal point process models.","keywords":["temporal point processes","event branches","Bregman ADMM","Hawkes process","self-attention","sparse low-rank regularization","expectation-maximization","interpretable transition matrices"],"falsifier":"A concrete test: simulate a Hawkes process with a known sparse branching matrix, run EM with and without the BADMM module over the full $\\lambda$/\\$\\alpha$ grid, and check whether (i) the log-likelihood of the BADMM run ever falls below plain EM on the same training data, and (ii) the recovered transition matrix matches the ground-truth branches (e.g., by precision/recall on nonzero entries). Failure on either count would undercut the claim that structured branch inference comes with improved performance.","tokens_in":1923,"feed_emoji":"🧩","tokens_out":3885,"duration_ms":85253,"temperature":0.7,"pith_summary":"Event sequences often hide a branching structure: which past event triggered which later event. This paper claims that the matrices TPP models already produce for this purpose—the responsibility matrix in EM for Hawkes processes and the attention map in transformer TPPs—can be turned into clean event branches by forcing them to be sparse and low-rank. The proposed BADMM module solves a regularized optimization problem with Bregman ADMM, and its unrolled iterations become a plug-in layer for neural models. Experiments on synthetic and real-world data report that plugging the module into Hawkes, THP, and SAHP improves log-likelihood and prediction accuracy compared with softmax and Sinkhorn baselines, while the learned matrices expose isolated events and key triggering events. The reader should care because this is a route to interpretability without giving up predictive performance.","feed_headline":"Sparse low-rank module sharpens event-branch inference","feed_subtitle":"Plugging it into Hawkes and transformer point-process models lifts accuracy and exposes triggering events","key_machinery":"The central object is the BADMM module: an iterative solver for the regularized KL-projection of a candidate transition matrix onto the set $\\Omega$ of row-normalized lower-triangular stochastic matrices. It alternates a closed-form update $$$B^{{(t+1)}}$=\\sigma_r\\Big(\\frac{\\log B_0+\\rho\\sum_{i=1}^2(\\log $X_i^{{(t)}}$-$Z_i^{{(t)}}$)}{1+2\\rho}\\Big)$$ with soft-thresholding updates for $X_1$ (the $\\ell_1$ term) and for $X_2$ (nuclear norm via SVD, or $\\ell_{1,2}$ via column-group thresholding), and dual ascent on $Z_1,Z_2$. Unrolling $T$ iterations gives a differentiable attention layer; for the SVD variant the gradient through $X_2$ is detached to save memory. This mechanism is what turns an over-smooth matrix into a sparse, low-rank branch structure.","core_discovery":"This paper's central claim is that the hidden branching structure behind an event sequence can be recovered by imposing sparse and low-rank structure on whatever matrix the model already uses to represent influence—the EM responsibility matrix for Hawkes processes or the attention map for transformer TPPs. Its BADMM module takes that initial matrix $B_0$ and solves $$\\min_{B\\in\\$\\Omega$} \\mathrm{KL}(B\\parallel B_0)+\\$\\lambda$(\\$\\alpha$\\|B\\|_1+(1-\\$\\alpha$)R(B)),$$ with $R$ either the nuclear norm (subspace clustering) or the $\\ell_{1,2}$ norm (sparse group-lasso). The solution, computed by alternating Bregman-ADMM updates and unrolled as network layers, becomes the learned event transition matrix. The paper's experiments show that plugging this module into HP, THP, and SAHP improves log-likelihood and prediction accuracy relative to softmax and Sinkhorn baselines, and that the resulting sparse branches identify isolated and triggering events; the 12 Angry Men case study ranks influential jurors more consistently with text-based judgments using only timestamps and speaker IDs.","pith_inferences":["Editorial inference: the method implicitly places a sparse-and-low-rank prior on branching structure; if formalized, it could be connected to Bayesian TPPs, and that view predicts the benefit shrinks when the true branching matrix is dense.","Editorial inference: the unrolled module with detached SVD gradient behaves like a two-level optimizer, so its training dynamics may differ from a pure differentiable layer; a study of gradient bias versus memory savings would be a natural extension.","Editorial inference: the module's reliance on row-normalized lower-triangular matrices assumes strict temporal ordering; adapting it to contemporaneous or cyclic triggering (e.g., mutual excitation in networked events) would be a testable extension."],"forward_implications":["For Hawkes processes trained by EM, the E-step can directly output a structured responsibility matrix, making isolated and key events visible without changing the closed-form M-step.","For transformer TPPs, replacing softmax attention with the unrolled BADMM layer yields attention maps that are simultaneously row-normalized, lower-triangular, sparse, and low-rank—properties softmax or Sinkhorn attention cannot jointly satisfy.","The two regularizer implementations give users a choice: nuclear-norm BADMM preserves local triggering patterns, while sparse-group-lasso BADMM identifies a few globally significant events.","Across Taobao, Retweet, StackOverflow, Amazon, and synthetic Conttime data, BADMM variants match or exceed softmax and Sinkhorn baselines on event-log-likelihood and type-prediction accuracy.","Event-branch insights transfer across domains: the 12 Angry Men study identifies top influential jurors from timestamps and speaker IDs alone, agreeing with text-based rankings."],"supporting_citations":[{"why":"Supplies the Bregman ADMM algorithm whose iterates the module unrolls into network layers.","marker":"(Wang and Banerjee 2014)"},{"why":"Provides the EM algorithm for Hawkes processes that the classical branch modifies in the E-step.","marker":"(Zhou, Zha, and Song 2013b)"},{"why":"Motivates the nuclear-norm low-rank implementation as subspace clustering.","marker":"(Elhamifar and Vidal 2013)"},{"why":"Motivates the $\\ell_{1,2}$ implementation as sparse group-lasso.","marker":"(Simon et al. 2013)"},{"why":"Defines the Transformer Hawkes process backbone whose attention map is replaced by the BADMM module.","marker":"(Zuo et al. 2020)"},{"why":"Defines the Self-Attentive Hawkes process backbone used as a second neural testbed.","marker":"(Zhang et al. 2020)"},{"why":"Provides the Sinkhorn-based attention baseline that the paper compares against for structured attention.","marker":"(Sander et al. 2022)"},{"why":"Provides HPST, the source-tracking baseline and the textual-conversation setting used in the 12 Angry Men evaluation.","marker":"(Zhang et al. 2018)"},{"why":"Precedent for unrolling optimization iterations as differentiable layers, which the BADMM module follows.","marker":"(Xu and Cheng 2023)"},{"why":"Justifies treating Hawkes process branches as a transition matrix and supplies the branching-process viewpoint the module operationalizes.","marker":"(Møller and Rasmussen 2006)"}],"fun_headline_variants":["Bregman ADMM unrolls to reveal event trigger branches","Sparse low-rank plug-in decodes event sequences","Plug-and-play Bregman module sharpens branch inference","Unrolled Bregman ADMM exposes causal event links","Sparse attention maps unravel event branching structure"],"cache_read_input_tokens":17408,"weakest_assumption_plain":"The load-bearing premise is that substituting the BADMM-regularized matrix for the EM responsibility matrix in the E-step still allows the EM algorithm to improve the observed-data likelihood; the paper uses this substitution for the Hawkes experiments without proving that the EM lower bound is preserved or that the iterates converge.","fun_headline_variants_meta":{"raw":{"variants":["Bregman ADMM unrolls to reveal event trigger branches","Sparse low-rank plug-in decodes event sequences","Plug-and-play Bregman module sharpens branch inference","Unrolled Bregman ADMM exposes causal event links","Sparse attention maps unravel event branching structure"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000886,"raw_usage":{"total_tokens":3890,"prompt_tokens":1076,"completion_tokens":2814,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":692,"completion_tokens_details":{"reasoning_tokens":2735}},"tokens_in":692,"tokens_out":2814,"duration_ms":19130,"temperature":1.0,"reasoning_tokens":2735,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:30:20.571647+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test: simulate a Hawkes process with a known sparse branching matrix, run EM with and without the BADMM module over the full $\\lambda$/\\$\\alpha$ grid, and check whether (i) the log-likelihood of the BADMM run ever falls below plain EM on the same training data, and (ii) the recovered transition matrix matches the ground-truth branches (e.g., by precision/recall on nonzero entries). Failure on either count would undercut the claim that structured branch inference comes with improved performance.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Motivates the $\\ell_{1,2}$ implementation as sparse group-lasso."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Bregman ADMM algorithm whose iterates the module unrolls into network layers."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Motivates the nuclear-norm low-rank implementation as subspace clustering."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Self-Attentive Hawkes process backbone used as a second neural testbed."},{"cited_title":"E.; Ablin, P.; Blondel, M.; and Peyr \\'e , G","cited_arxiv_id":null,"evidence_quote":"Provides the Sinkhorn-based attention baseline that the paper compares against for structured attention."},{"cited_title":"Who Started It? Identifying Root Sources in Textual Conversation Threads","cited_arxiv_id":"1809.03648","evidence_quote":"Provides HPST, the source-tracking baseline and the textual-conversation setting used in the 12 Angry Men evaluation."}],"review_version":1}