{"id":"4ad44397-e394-4214-a259-7aae89c00e54","arxiv_id":"1908.05976","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A static graph layout that adds forces from higher-order causal path models to expose temporal clusters and temporal closeness in dynamic networks.","lead":"HOTVis is a new graph-drawing algorithm that uses higher-order models of temporal paths to make static layouts of dynamic networks, pulling closer the nodes that can influence each other over time. It matters for exploratory analysis of time-stamped data, from social interactions to transport, because it reveals temporal clusters and central nodes that ordinary static layouts miss.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The evaluation relies on quality measures σ and Δ that directly reward the same endpoint-pulling forces HOTVis is built to apply, so the empirical support for the causal-topology claim is not yet established; a quantitative timestamp-shuffle control would settle whether the ordering of edges…","rationale":"The paper makes a clear, falsifiable claim: static layouts can encode causal-path topology. The algorithm is specified with pseudocode, the α_k normalisation is parameter-free, code and data are released, and the held-out-path evaluation is a genuine attempt to test generalisation. Those are real strengths. My concern is not that the algorithm is wrong but that the quantitative evidence is not yet independent of the construction. σ and Δ are mathematical restatements of what the force term optimises, so the observed improvements could occur even if visualising causal topology were useless to a viewer. The timestamp-shuffle control is the natural experiment: it preserves edge frequencies and the number of forces per order (because α_k is renormalised by the number of unique paths) while destroying the causal ordering. The paper performs this shuffle only in a qualitative synthetic figure. Repeating it quantitatively on the empirical data would either validate that the ordering of edges drives the σ/Δ changes or show that the measures are circular. I therefore retain the reader's CONDITIONAL verdict: the central idea is plausible and reproducible, but the empirical support should be strengthened before the claim is accepted as established.","tokens_in":12226,"tokens_out":7490,"duration_ms":81581,"concrete_test":"On each of the five empirical datasets, recompute the Figure 3 σ curves and Figure 5 Δ curves (including the ROC-AUC prediction of top-10% temporal closeness) for K=1 through Kopt+2 after randomly permuting the time stamps of all edges, using identical held-out splits and the same αk=m_k^{-1} normalisation. If the shuffled condition produces σ reductions and AUC gains comparable to the unshuffled condition, the reported improvements are driven by the added force budget rather than by causal order; if the shuffled curves revert to the K=1 level, the causal-topology explanation is corroborated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Algorithm 1 (lines 4–10), HOTVis adds attractive forces between the endpoints of every causal path of length k. In Section 4, the causal-path dispersion σ measures precisely whether the vertices of causal paths are less spatially dispersed than random paths, and the closeness eccentricity Δ measures whether high-temporal-closeness vertices sit near the barycentre. Both metrics therefore encode the algorithm's own objective rather than an independent criterion of visual quality or readability. The held-out split only shows that the effect generalises to unseen causal paths; it does not break the circularity, because the layout was still optimised to pull endpoints of training paths together and the test paths share the same temporal structure. The only baseline is the K=1 first-order layout, so there is no evidence relative to existing dynamic-graph visualisation methods (time-unfolded DAGs, time-sliced layouts, space-time cubes). The synthetic timestamp-shuffle panel is qualitative only: σ and Δ are never computed on shuffled data, and no quantitative test distinguishes \"the temporal ordering of edges is what matters\" from \"any set of length-2 paths, causal or not, gets pulled together by the extra forces.\" The school-data cluster demonstration is visual inspection only, without a layout-clustering metric against ground-truth classes. Because these measures are used in all of the paper's quantitative claims, this is the load-bearing weakness.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HOTVis, a static layout algorithm for dynamic graphs. It builds higher-order De Bruijn graph models of causal paths up to a maximum order K, and generalises the Fruchterman-Reingold force-directed layout by adding attractive forces between the endpoints of each causal path in addition to first-order edges (Algorithm 1). The authors claim that the resulting layouts highlight temporal clusters that are not visible in time-aggregated layouts and place vertices with high temporal closeness centrality near the centre. They evaluate on one synthetic model and five empirical data sets (flights, tube, hospital, workplace, school) using edge crossings, a custom causal-path dispersion σ, and a custom closeness eccentricity Δ, with 100-fold held-out cross-validation and statistical model selection for K. The paper is accompanied by open-source code and data.","tokens_in":12519,"tokens_out":6100,"duration_ms":64671,"significance":"If the evaluation is accepted, the contribution fills a real gap in temporal-network visualisation: static layouts that encode causal topology rather than only aggregated topology. The algorithmic idea is simple and plausible, the implementation is public, and the cross-validation protocol (100 runs, held-out paths) is a strength. The timestamp-shuffle demonstration in Figure 2, although qualitative, gives some evidence that the effect is due to temporal ordering. However, the paper's quantitative quality measures are defined over exactly the causal paths whose endpoints the algorithm is designed to pull together, and the only baseline is a first-order layout. Independent evaluation is needed before the central claim can be considered established.","major_comments":[{"comment":"The causal path dispersion σ and closeness eccentricity Δ are aligned with the objective of Algorithm 1 (lines 7–10), which adds attractive forces between the endpoints of every causal path. A lower σ or Δ for K > 1 therefore partly reflects the algorithm's own optimisation target rather than an independent property of visual quality. The held-out split does not resolve this, since the validation paths share the same temporal structure as the training paths. Please add an independent criterion, e.g., a task-based user study, a readability metric from the graph-drawing literature, a quantitative timestamp-shuffle control in which σ and Δ are compared on original versus shuffled timestamps, or a comparison against existing dynamic-graph drawing methods.","section":"Section 4 (definitions of σ and Δ)"},{"comment":"The only baseline in the experimental evaluation is K = 1, i.e., the Fruchterman-Reingold layout. The paper does not compare against existing static representations of dynamic graphs, such as time-unfolded directed acyclic graphs, time-sliced layouts, space-time cubes, or the higher-order visualisation of [35]. As a result, the claim that HOTVis provides better visual representations of causal topology is not supported relative to the state of the art. The authors should state which existing methods are applicable to high-resolution time-stamped data and, if possible, compare quantitatively on the same data sets.","section":"Section 4 (baselines)"},{"comment":"The claim that HOTVis reveals ground-truth class structure in the school data is based on visual inspection only. The left and right panels are not accompanied by a quantitative comparison, such as a clustering metric between spatial positions and class labels or an inter-class versus intra-class causal-path dispersion measure. Without such a measure, the temporal-cluster demonstration on empirical data remains anecdotal rather than established.","section":"Section 4 (Figure 4, school data)"}],"minor_comments":[{"comment":"The statement that σ ≈ 1 corresponds to the same dispersion as 'random paths' is not exact: the numerator averages distances to each path's own barycentre, while the denominator averages distances to the global barycentre, so for random paths of length l the expected ratio is approximately 1 − 1/l. The normalization should be clarified or replaced by an explicit random-path baseline.","section":"Section 2 (causal path dispersion definition)"},{"comment":"The definition of the weight w in Section 2 uses the notation `{|...|}`; this should be written as a cardinality, and the indexing of the timestamps in the causal-path definition should be stated more precisely.","section":"Section 2 (higher-order graph weights)"},{"comment":"In the synthetic model, steps 3–4 introduce t1 and t2 without defining their ranges; clarify that the time stamps are chosen so that the swap preserves the multiset of time-stamped edges.","section":"Appendix G"},{"comment":"The parameter δ is passed to HigherOrderGraph but is not defined in the pseudocode; state in the caption or text that it is the maximum time difference from Section 2.","section":"Algorithm 1"},{"comment":"There are minor typos, e.g., 'we now tests whether' and inconsistent subject-verb agreement in the discussion of σ. A careful proofread would improve clarity.","section":"Section 4.1"},{"comment":"The related-work section mentions HonVis [35] but does not explain how HOTVis differs in output representation from that system; a sentence describing the difference would help position the contribution.","section":"Related Work"}],"recommendation":"major_revision","confidential_remarks":"I see this as a genuinely interesting but incrementally evaluated contribution. The core algorithm is simple and the topic is timely. My main reservation is that the evaluation measures success with quantities that the algorithm is explicitly built to optimise; this is fixable with a stronger control and at least one external baseline. I would support a major revision rather than a rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a real, small contribution. HOTVis generalizes force-directed layouts by adding attractive forces between endpoints of causal paths from higher-order De Bruijn models, so temporal ordering leaves a trace in a static layout. That is precisely the hole Holme pointed to, and I do not find this combination in the cited prior work.\n\nWhat the paper does well: Algorithm 1 is concrete and clearly specified; code and data are public; the evaluation uses held-out causal paths, 100 runs, and a timestamp-shuffle control. The strongest piece is the ROC experiment: layout proximity predicts temporal closeness centrality computed from data, and the AUC improves over the K=1 baseline in all five empirical datasets. That is independent evidence the layout carries causal-topology information, because the algorithm is not directly optimizing barycentric closeness for high-centrality nodes. The authors also report the edge-crossing increase rather than hiding it.\n\nThe soft spots are real but not fatal. The custom measures σ and Δ are closely aligned with the algorithm's objective: Algorithm 1 pulls endpoints of causal paths together, and σ measures whether causal-path vertices are spatially less dispersed. Showing σ drops with K is partly a self-consistency check, not an independent quality test. The held-out split only shows the effect generalizes to unseen paths with the same temporal structure. The timestamp-shuffle control is qualitative only—they never report σ or Δ on shuffled data—so it does not quantitatively separate \"temporal ordering matters\" from \"any length-2 paths get pulled together.\" And there is no comparison against existing dynamic graph drawing baselines such as time-unfolded DAGs, time-sliced layouts, or space-time cubes. The school-data cluster separation is visual inspection only, though the ROC analysis partially compensates.\n\nIf I were refereeing, I would not desk-reject this. The method is sound, the gap is real, and the public artifacts earn credit. But the central claim \"better represents causal topology\" needs either a non-aligned readability metric or a proper baseline comparison. The timestamp-shuffle control should be made quantitative. Those are reasonable revision requests, not grounds for rejection.\n\nWho is this for? Researchers in temporal network visualization and anyone doing data mining on time-stamped interaction data. I would cite it. It deserves a serious peer review, with the revision items above.","headline":"HOTVis is a genuine, modest step forward for temporal network visualization, but its central empirical claim is only partly established because the quality metrics largely reward exactly what the algorithm optimizes.","tokens_in":12993,"tokens_out":2634,"would_cite":true,"duration_ms":30029,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68R10","05C82","68U05"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper presents HOTVis, a force-directed layout algorithm that adds attractive forces along causal paths in time-stamped networks, producing static visualisations in which temporal clusters and temporally central vertices become…","keywords":["dynamic graph visualisation","temporal networks","causal paths","higher-order graph models","force-directed layout","temporal clusters","closeness centrality","static time-aware layout"],"falsifier":"On a synthetic temporal network with known ground-truth temporal clusters, apply a standard spatial clustering algorithm to the 2D positions produced by HOTVis and compare cluster recovery for $K=1$ versus $K>1$, for example with the adjusted Rand index between recovered and true clusters. If recovery does not improve for $K>1$, the claim that HOTVis makes temporal clusters visible in static layouts is not supported.","tokens_in":12032,"feed_emoji":"🕸️","tokens_out":12817,"duration_ms":108675,"temperature":0.7,"pith_summary":"This paper sets out to fix a blind spot in dynamic graph drawing: standard static visualisations collapse time-stamped edges into an aggregated graph and discard the ordering information that determines causal paths. It presents HOTVis, a layout algorithm that generalises force-directed drawing by adding attractive forces between the endpoints of causal paths, extracted from higher-order De Bruijn graph models of the temporal data. The intended payoff is a single static picture that reveals temporal clusters and temporally central vertices, structures that exist only because of edge ordering and are invisible in time-aggregated layouts. The paper reports experiments on a synthetic temporal network with known clusters and on five empirical data sets (flight itineraries, London Underground trips, and three face-to-face contact networks), finding lower causal-path dispersion and stronger central placement of temporally close vertices for higher orders.","feed_headline":"Causal-path forces reveal temporal clusters hidden in static graphs","feed_subtitle":"HOTVis adds attractive forces along time-respecting paths, turning dynamic edge order into a static picture of influence.","key_machinery":"The load-bearing object is the higher-order De Bruijn graph $G^{(k)}$ of causal paths: its vertices are sequences of $k$ vertices that occur as time-respecting paths in the dynamic graph, and its edges connect overlapping sequences, so an edge $(\\overrightarrow{v_0\\ldots v_{k-1}},\\overrightarrow{v_1\\ldots v_k})$ encodes a causal path $v_0\\to\\cdots\\to v_k$ that actually occurs in the data. HOTVis superimposes on the standard force-directed many-body simulation an attractive force between $v_0$ and $v_k$ for each such higher-order edge, with strength proportional to the path frequency and normalised across orders by $\\alpha_k = 1/m_k$, where $m_k$ is the number of unique paths of length $k$. The first-order case $\\alpha_k=0$ for $k>1$ reproduces the time-aggregated force-directed layout, so the higher-order attractive forces are the mechanism that carries the temporal information into the drawing.","core_discovery":"On the paper's own terms, the central claim is that the causal topology of a dynamic graph—which nodes can influence which other nodes through time-respecting paths—can be rendered visible in a static layout if the attractive forces are computed from causal paths rather than from static edges. HOTVis builds $k$-th order De Bruijn graph models of causal paths up to a maximum order $K$; for every higher-order edge $(\\overrightarrow{v_0\\ldots v_{k-1}},\\overrightarrow{v_1\\ldots v_k})$ it adds an attractive force between $v_0$ and $v_k$, scaled by path frequency and by an order weight $\\alpha_k$. With $\\alpha_k=0$ for all $k>1$ the algorithm reduces to the standard force-directed layout, so the time-aware drawing is a strict generalisation of the first-order baseline. In the reported experiments, increasing $K$ reduces the causal-path dispersion $\\sigma$ and the closeness eccentricity $\\Delta$ across all data sets, and a simple rule that ranks nodes by distance to the barycentre predicts top-10% temporal closeness with higher area-under-the-curve scores for $K>1$ than for $K=1$. The paper reads these results as evidence that HOTVis highlights temporal clusters and places temporally central vertices near the centre of the drawing.","pith_inferences":["The two quality measures $\\sigma$ and $\\Delta$ reward exactly the proximity HOTVis creates, so a stronger test would be to cluster the 2D positions and compare the recovered clusters with ground-truth temporal clusters; the paper does not run that test.","The same superposition idea could be ported to other layout families, such as spectral or multidimensional-scaling layouts, to test whether causal topology can be encoded without the many-body force simulation.","The choice of $K$ could be driven by visual readability as well as statistical model selection, which would turn HOTVis into an interactive tool that adapts the drawing to the user's task.","The area-under-the-curve result suggests a practical shortcut for monitoring: use distance to the barycentre as a real-time proxy for temporal closeness centrality, at least for data whose causal structure resembles the tested networks."],"forward_implications":["Temporal clusters that exist only in the ordering of edges, not in the aggregated topology, can be read directly from a single static image.","High-resolution dynamic data can be visualised without time slicing or animation, because the method uses exact time stamps to define causal paths.","The maximum order $K$ can be selected by statistical model selection rather than manual tuning, giving a principled trade-off between causal-topology fidelity and layout clutter.","Vertices with high temporal closeness become visually identifiable by their central position, so exploratory ranking of important nodes no longer requires a separate computation.","The same layout principle applies across data types that produce causal paths, from passenger itineraries to face-to-face contact sequences, and stays practical for graphs with millions of time-stamped edges."],"supporting_citations":[{"why":"supplies the many-body simulation that HOTVis generalises and the first-order layout baseline it must beat.","marker":"[12]"},{"why":"provides the line-graph construction used to build higher-order vertices and edges from causal paths.","marker":"[5]"},{"why":"defines time-respecting causal paths and motivates the need for static visualisations that preserve them.","marker":"[16]"},{"why":"establishes higher-order network models as the representation of causal topology in time-stamped data.","marker":"[21]"},{"why":"supplies the multi-order graphical model selection method that determines the optimal maximum order Kopt.","marker":"[31]"},{"why":"supplies the US flight itinerary data used in the empirical evaluation.","marker":"[36]"},{"why":"supplies the London Underground trip data used in the empirical evaluation.","marker":"[22]"},{"why":"supplies the hospital proximity-contact data set used in the evaluation.","marker":"[37]"},{"why":"supplies the primary-school face-to-face contact data set used in the evaluation.","marker":"[34]"},{"why":"supplies the office workplace face-to-face contact data set used in the evaluation.","marker":"[13]"}],"fun_headline_variants":["Causal-path forces reveal temporal clusters in static layouts","HOTVis: higher-order paths drive time-aware static graph drawing","Path-based forces make dynamic causality visible in static graphs","Static graph layouts now encode temporal influence via causal paths","Causal topology drawn with higher-order attractive forces"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes that the two custom quality measures—how tightly causal paths are packed and how central the temporally close nodes sit—are the right criteria for a good visualisation, and both measures reward exactly the proximity that HOTVis is built to create.","fun_headline_variants_meta":{"raw":{"variants":["Causal-path forces reveal temporal clusters in static layouts","HOTVis: higher-order paths drive time-aware static graph drawing","Path-based forces make dynamic causality visible in static graphs","Static graph layouts now encode temporal influence via causal paths","Causal topology drawn with higher-order attractive forces"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000241,"raw_usage":{"total_tokens":1541,"prompt_tokens":983,"completion_tokens":558,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":480}},"tokens_in":599,"tokens_out":558,"duration_ms":6409,"temperature":1.0,"reasoning_tokens":480,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:58:49.455206+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a synthetic temporal network with known ground-truth temporal clusters, apply a standard spatial clustering algorithm to the 2D positions produced by HOTVis and compare cluster recovery for $K=1$ versus $K>1$, for example with the adjusted Rand index between recovered and true clusters. If recovery does not improve for $K>1$, the claim that HOTVis makes temporal clusters visible in static layouts is not supported.","supporting_citations":[{"cited_title":"Software: Practice and experience21(11), 1129–1164 (1991)","cited_arxiv_id":null,"evidence_quote":"supplies the many-body simulation that HOTVis generalises and the first-order layout baseline it must beat."},{"cited_title":"Koninklijke Nederlandse Akademie v","cited_arxiv_id":null,"evidence_quote":"provides the line-graph construction used to build higher-order vertices and edges from causal paths."},{"cited_title":"The European Physical Journal B 88(9), 234 (2015)","cited_arxiv_id":null,"evidence_quote":"defines time-respecting causal paths and motivates the need for static visualisations that preserve them."},{"cited_title":"Nature physics15(4), 313–320 (2019)","cited_arxiv_id":null,"evidence_quote":"establishes higher-order network models as the representation of causal topology in time-stamped data."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the multi-order graphical model selection method that determines the optimal maximum order Kopt."},{"cited_title":"transtats.bts.gov/Tables.asp?DB_ID=125","cited_arxiv_id":null,"evidence_quote":"supplies the US flight itinerary data used in the empirical evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the London Underground trip data used in the empirical evaluation."},{"cited_title":"PloS one8(9), e73970 (2013)","cited_arxiv_id":null,"evidence_quote":"supplies the hospital proximity-contact data set used in the evaluation."},{"cited_title":"PloS one6(8), e23176 (2011)","cited_arxiv_id":null,"evidence_quote":"supplies the primary-school face-to-face contact data set used in the evaluation."},{"cited_title":"Network Science3(3), 326–347 (2015)","cited_arxiv_id":null,"evidence_quote":"supplies the office workplace face-to-face contact data set used in the evaluation."}],"review_version":1}