{"id":"cd9ce4cb-13a3-47c0-87b7-a8514dad9023","arxiv_id":"2607.24506","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Semi-relaxed Fused Gromov–Wasserstein partitioning of attributed transportation graphs gives explicit α-control over structure versus heterogeneous attributes, shown on a French road network and London bike-share.","lead":"The paper gives a way to split transportation networks into zones by balancing road layout against traffic or usage data, using optimal transport. Practitioners can dial how much topology versus attributes matter, which existing methods rarely allow cleanly.","discovery_kind":"new_application","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The reported normalized-total-variance advantage is not yet independent evidence: srFGW directly minimizes distance-to-barycenter variance, while α and k are selected using the same TVn-based evaluation.","rationale":"The paper is clear that srFGW is intended as a flexible simultaneous-fusion method rather than a universally better clustering algorithm. Its central trade-off mechanism is internally coherent, and the two case studies provide plausible qualitative evidence that α changes the partition structure. The soft spot is specifically the benchmark interpretation. Because the method’s barycentric attribute objective is closely related to TVn, superior in-sample TVn is expected unless the evaluation is separated from model fitting and hyperparameter selection. The same issue is less direct for silhouette and modularity in the bicycle case, but attribute-distance-to-barycenter optimization can still favor silhouette-like separation.\n\nI therefore agree with the Reader’s weakest assumption and do not see a reason to change the CONDITIONAL verdict. The contribution remains accept-shaped if the claims are phrased primarily as explicit, adaptable fusion and if superiority is either validated out of sample or explicitly described as performance under the task-selected criteria. The proposed temporal holdout is especially appropriate because it tests whether the learned partitions capture recurring traffic organization rather than merely minimizing variance in the same attribute measurements used for fitting and tuning.","tokens_in":23188,"tokens_out":1786,"duration_ms":108117,"concrete_test":"Split the road FCD into disjoint temporal training and validation periods. For every method, build attribute distances from the training period, tune all hyperparameters using only training-period connectivity/TVn, and freeze the resulting node partition. Then recompute each node’s validation-period attributes/distances and evaluate that frozen partition’s TVn and connectivity. Repeat over several time splits and use common k values where applicable. If srFGW’s TVn advantage over Louvain/Snake-spectral largely disappears out of sample, the reported gain is substantially objective- and selection-driven; if it persists, the comparative claim is materially strengthened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The road-network claim of superior normalized total variance is vulnerable to objective/evaluation alignment. In Eq. (1), srFGW explicitly includes the squared attribute distance from each source node to a class barycenter, and Algorithm 1 alternately updates those barycenters and assignments. The evaluation in §4.1.3 then minimizes TVn, which is also within-cluster attribute variance, and §4.2.2 uses a normalized combination of connectivity and \\(1-TVn\\) to select \\(\\alpha\\) and \\(k\\). Louvain and Snake-spectral clustering instead optimize modularity/Snake-similarity objectives and are then scored on a criterion they were not designed to minimize. Consequently, the statement that srFGW performs “by far the best” on TVn may partly measure alignment with its own objective and hyperparameter-selection rule rather than superior fusion. This does not undermine the narrower flexibility claim—the \\(\\alpha\\)-sweeps do show controllable structure–attribute trade-offs—but it weakens the comparative superiority component of the strongest claim.","agreement_with_reader":"agree"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The paper proposes a framework for partitioning attributed transportation graphs using the semi-relaxed Fused Gromov–Wasserstein (srFGW) discrepancy. Nodes (road segments, bike stations) are represented via heterogeneous attribute distance matrices (DTW on traffic curves, Wasserstein-1 on speed/duration histograms, Euclidean on scalars), and partitioning is cast as optimal transport from the source graph to a fixed k-node target graph with uniform inter-cluster distance mean(D_S), alternating between solving the srFGW problem and updating cluster attribute barycenters (Algorithm 1). The parameter α gives explicit control of the structure–attribute trade-off, which is the central claim. The method is demonstrated on a small French road network (Châteaubourg, dual graph of street segments with FCD attributes) and on the London Cycle Hire origin–destination network, compared against Louvain on a weighted adjacency and spectral clustering on Snake similarity.","tokens_in":23473,"tokens_out":3154,"duration_ms":112921,"significance":"If the results hold, the contribution is a practically useful, well-motivated application of srFGW to transportation partitioning: α gives an interpretable knob for structure–attribute fusion, and the two case studies on genuinely different network types (a primal/dual road graph and a flow-derived station graph) do demonstrate that the partitions move controllably as α and k vary (Figures 6–8, 13). The paper ships real strengths: publicly available code (GitHub, built on POT) and data (Zenodo, TfL open data), 100-repetition runs for all stochastic methods, and honest acknowledgment of limitations (no connectivity guarantee, scalability). The interpretive results — e.g., recovering a major/medium/minor road typology from normalized attributes without volume information (Table 3), and splitting the LCH west zone into functionally distinct communities — are genuinely informative. The work is incremental over [67] (same srFGW-for-clustering machinery, by overlapping authors) but the transportation-specific representation layer (Hausdorff-variant handling of two-directional attributes, distance-based unification of heterogeneous attributes) is a real contribution to this venue's reades","major_comments":[{"comment":"§3.3 (Speed Histograms): the displayed formula for W_1 is incorrect as written. W_1 between one-dimensional histograms over ordered bins equals (up to normalization) the L1 distance between the cumulative distribution functions, sum_t |F_i(t) - F_j(t)|. The formula given, (1/T) sum_t |h^(i)_t - h^(j)_t|, is a normalized L1/total-variation distance between the histograms themselves and is not shift-sensitive in the way the surrounding text claims ('sensitive to shifts, captures differences in shape'). If the implementation actually uses cumulative sums (e.g., via POT's wasserstein_1d), the formula must be corrected; if it uses the displayed formula, the distance should be renamed and the interpretation revised. This is load-bearing because the Wasserstein framing is part of the paper's claim to handle distributional attributes properly.","section":"§3.3"},{"comment":"§4.1.3, definition of TVn: 'var(C_i) denotes the variance of the attributes within cluster C_i' is underspecified. The attributes are heterogeneous — traffic curves, speed histograms, scalars — and no natural scalar 'variance of the attributes' exists. Presumably this is a length-weighted average of squared attribute distances to a cluster centroid (possibly computed per attribute distance matrix), which would make TVn essentially a normalized version of the srFGW attribute-fidelity term in Eq. (1). The exact formula must be stated, including which distance(s) enter, whether they are the min–max normalized and square-root-transformed distances of §3.3, and how the cluster reference point is defined. Without this, the paper's headline road-case result (srFGW 'by far the best' on TVn, Figure 6) cannot be assessed independently of the method's own objective.","section":"§4.1.3"},{"comment":"§4.2.2 and §5: the comparative claims rest partly on objective–evaluation alignment. srFGW directly minimizes within-cluster attribute distance to learned barycenters (Eq. 1), α and k are then selected by optimizing the combined connectivity / (1 - TVn) criterion (Figure 8), and the same TVn is the criterion on which srFGW is declared 'by far the best', while Louvain and Snake-spectral are scored on a criterion their objectives were not designed for. The α-sweeps do establish the narrower flexibility claim, which I find convincing; but the superiority claim needs either (a) a control in which baselines' own free parameters (resolution, adjacency-α, Snake prop) are selected by the same combined criterion and the gap is re-examined — Figure 6 partially does this but the selected operating points in Figure 9 mix parameter regimes — or (b) a held-out evaluation (e.g., attribute variance on d","section":"§4.2.2"},{"comment":"§5.2.1, modularity evaluation for the LCH case: the modularity formula is stated in terms of 'the adjacency matrix A_ij', but the LCH network is described with edge lengths equal to the inverse of trip counts (2454 edges after the 200-trip threshold). It is not stated whether modularity is computed on the unweighted or weighted adjacency, and whether the degrees k_i and m are weighted. Since modularity is one of the two criteria used to select α = 0.8 (Figure 13), this needs to be made explicit; the choice materially changes the criterion.","section":"§5.2.1"},{"comment":"Algorithm 1 and §4.1.2: no convergence analysis or diagnostics are reported for the alternating scheme (srFGW solve with fixed barycenters, then barycenter update from the soft plan, then hard assignment with the tie-breaking rule T_il > T_im or (T_il = T_im and l ≤ m)). The stopping test T_n = T_{n-1} may never trigger exactly for a non-convex solver, and the deterministic tie-breaking systematically favors low-index clusters, which can interact with empty-cluster handling (k redefined as the number of nonempty clusters each iteration). Please report observed iteration counts, objective trajectories across the 100 repetitions, and how often the maximum-iteration cap binds; a brief argument (or citation to [67]) that each step non-increases a common surrogate would strengthen the section.","section":"§4.1.2"}],"minor_comments":[{"comment":"§3.3: the square-root transformation of distance matrices and β = 1/2 in d_A are justified only by 'preventing trivial clusterings'; a brief sensitivity check (e.g., β ∈ {1/4, 1/2, 3/4} in the LCH or road case) would make the equal-weighting choice less arbitrary.","section":"§3.3"},{"comment":"§4.2.2, Figure 6: the caption states k = 6 and resolution = 0.04, but the Louvain curve varies with α; since Louvain's resolution was 'adjusted to produce a comparable range', state explicitly which resolution yields k = 6 at each α, or clarify that the plotted Louvain curve is at fixed resolution with varying cluster count.","section":"§4.2.2"},{"comment":"§4.1.3: 'classical criteria can also be employed like.' — sentence ends abruptly after 'like'; the intended criteria (modularity, silhouette) appear after the comma. Please fix the grammar.","section":"§4.1.3"},{"comment":"§2.2: 'limited limited by their recursive clustering process' — duplicated word. Also 'the performance of this family of method' (missing 's'), §3.1 'approximatively' → 'approximately'.","section":"§2.2"},{"comment":"Figure 9 and Figure 10: cluster colors are overloaded across figures and text (e.g., 'green', 'cyan', 'yellow' refer to different entities in Figures 9, 10, 14); a legend keyed to each map and consistent naming between figure and prose would help readability.","section":"Figures"},{"comment":"§5.2.1: the 200-trip edge threshold and the exclusion of isolated stations/components are reasonable but their impact on the resulting communities is not quantified; one sentence on how many stations/trips are discarded and whether results are stable to, say, a 100-trip threshold would be useful.","section":"§5.2.1"},{"comment":"Related work: the srFGW clustering methodology closely follows [67] (same first-author group, 'under review'); please make the delta from [67] explicit in the introduction — as written, a reader must infer that the novelty here is the transportation representation layer and the two case studies, not the OT formulation itself.","section":"§1"}],"recommendation":"minor_revision","confidential_remarks":"The core srFGW partitioning machinery is from the authors' own under-review preprint [67] (arXiv:2512.15570), which is cited appropriately; the incremental contribution here is the transportation-specific attributed-graph construction and the two case studies. That is a legitimate scope for this venue, but the editor may wish to verify that [67]'s venue/status does not create substantial overlap. The main technical repairs (W_1 formula, TVn definition, modularity weighting, tempered comparative language) are all achievable without new experiments of any scale."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a clean transportation-methods paper that packages distance-based attributed graphs with semi-relaxed Fused Gromov–Wasserstein so you can dial structure vs attributes with α. The math engine is not new (Vayer/Vincent-Cuaz, plus their concurrent arXiv:2512.15570); what is new is the full transport-facing stack—DTW, Wasserstein histograms, Hausdorff for two-way streets, dual road graphs, usage networks—and two real case studies with Louvain and Snake-spectral baselines.\n\nWhat it does well: the representation section is careful and useful. They show that early fusion (weighted adjacency, combined distances, Snake length) already changes partitions a lot, then give a simultaneous-fusion alternative with an explicit trade-off. The α sweeps actually move connectivity and attribute homogeneity in the expected directions. The London bike case is more convincing than the road one for community-style reading: park/periphery self-loop cluster, residential vs destination patterns, weekday/weekend profiles in one pass. Criteria match the claimed goals, baselines are standard, and they ship code/data intent.\n\nSoft spots, in proportion. The stress-test lands partly: srFGW’s attribute term is distance-to-barycenter, Algorithm 1 updates those barycenters, and they pick α/k with a score that includes 1−TVn. Saying srFGW is “by far the best” on TVn is partly objective alignment, not pure external proof of better fusion. Louvain was never optimizing TVn. That weakens comparative superiority, not the narrower flexibility claim, which the sweeps still support. Mild metric-tuned hyperparameter circularity is normal in clustering; they should just say it. Road partitions do not enforce connectivity (they admit it); for MFD-style zoning that matters. Graphs are modest. Novelty is applied, not a new discrepancy.\n\nWho it’s for: people who partition roads or mobility graphs and want heterogeneous attributes without baking the mix into edge weights once and for all. Not for OT theorists hunting a new bound.\n\nI’d send it to peer review. Scoped claims, honest experiments, reproducible intent. Ask referees to push external task metrics and to tone the TVn superiority language. Worth engaging if you work in this lane.","headline":"Solid applied transport methods paper: usable α-controlled srFGW partitioning on real attributed networks, incremental on the OT side, with a fair but overstated TVn win.","tokens_in":24200,"tokens_out":570,"would_cite":true,"duration_ms":20546,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A single tunable optimal-transport discrepancy can partition transportation networks by structure and attributes together, instead of baking their mix into the graph up front.","keywords":["transportation networks","network partitioning","attributed graphs","optimal transport","Fused Gromov–Wasserstein","community detection","traffic clustering","bicycle-sharing systems"],"falsifier":"On the same road and bike-share instances, if systematically varying the target inter-cluster distance, initialization, or hard-assignment rule erases srFGW’s reported gains in normalized total variance (or modularity/silhouette) relative to Louvain and Snake spectral clustering at matched structure–attribute balance, the claimed advantage of joint OT fusion would not hold.","tokens_in":24057,"feed_emoji":"🛣️","tokens_out":928,"duration_ms":22317,"temperature":0.7,"pith_summary":"Transportation networks are graphs that carry both topology (who connects to whom) and messy operational attributes (speeds, traffic curves, trip-duration histograms, usage intensity). Most partitioning methods fuse those two sources early—into edge weights, a similarity matrix, or a fixed “Snake” representation—so the balance between structure and attributes is hard to dial. This paper argues that representing both structure and attributes as distance matrices, then partitioning with the semi-relaxed Fused Gromov–Wasserstein discrepancy, lets the user control that trade-off with one parameter while jointly optimizing both. On a small French road network the method recovers functional road types and traffic regimes; on London’s bike-share system it recovers usage communities that mix geography with arrival/departure profiles and self-loops. The claim is not a new universal clustering theorem, but a practical, adaptable pipeline: distance-based attributed graphs plus srFGW give partitions that bend toward connectivity or attribute homogeneity as the application demands.","feed_headline":"One transport dial balances roads’ shape and traffic data","feed_subtitle":"Semi-relaxed Fused Gromov–Wasserstein partitions cities and bike networks without locking the mix early","key_machinery":"Semi-relaxed Fused Gromov–Wasserstein (srFGW): an optimal-transport plan that matches the source network to a small target of k super-nodes by minimizing a convex mix of attribute transport cost and structural (Gromov) distortion, with the second marginal relaxed so cluster sizes need not be fixed in advance.","core_discovery":"When transportation systems are cast as attributed graphs with separate structural and attribute distance matrices, the semi-relaxed Fused Gromov–Wasserstein discrepancy yields node partitions that jointly respect topology and heterogeneous attributes, with an explicit scalar controlling their relative weight—something early-fusion methods (weighted adjacency, Snake similarity) do not provide as cleanly.","pith_inferences":["The same distance-plus-srFGW template could extend to multimodal networks (bus+rail+bike) if each mode supplies its own attribute distance block and a shared or layered structural distance.","Adding an explicit connectivity penalty inside the OT objective, as the authors flag as future work, would make the method competitive for parallel simulation partitions where cut size and balance dominate.","Because GW matches relational roles, the framework may surface “structurally analogous but distant” components (e.g., multiple park-edge bike hubs) that local modularity methods systematically miss."],"forward_implications":["Traffic managers can tune one parameter toward connected regions for control strategies or toward attribute-homogeneous clusters for Macroscopic Fundamental Diagram analysis.","Bike-share and other OD-based systems can cluster stations by usage profiles and network role in one run, without separate weekday/weekend analyses.","Scalar, curve, and histogram attributes can sit in one pipeline via pairwise distances (DTW, Wasserstein-1, Euclidean) without forcing a single early fusion formula.","Cluster barycenters produced by the OT plan give interpretable prototype roads or stations for each community."],"fun_headline_variants":["Optimal transport dial joins road topology with traffic attributes","Semi-relaxed FGW partitions cities by structure and ops data","One scalar steers network cuts across shape and usage profiles","Attributed graphs split via fused Gromov-Wasserstein trade-off","Transport networks yield tunable partitions of form and flow"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The method assumes that a fixed simple target geometry, uniform node masses, and a k-means-started transport plan produce fair partitions whose quality the same connectivity or modularity scores used to pick the trade-off parameter can honestly measure.","fun_headline_variants_meta":{"raw":{"variants":["Optimal transport dial joins road topology with traffic attributes","Semi-relaxed FGW partitions cities by structure and ops data","One scalar steers network cuts across shape and usage profiles","Attributed graphs split via fused Gromov-Wasserstein trade-off","Transport networks yield tunable partitions of form and flow"]},"model":"grok-4.5","effort":"low","cost_usd":0.003605,"raw_usage":{"total_tokens":1138,"prompt_tokens":697,"num_sources_used":0,"completion_tokens":83,"cost_in_usd_ticks":36048000,"prompt_tokens_details":{"text_tokens":697,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":358,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":697,"tokens_out":83,"duration_ms":5939,"temperature":1.0,"reasoning_tokens":358,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T12:54:29.965702+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On the same road and bike-share instances, if systematically varying the target inter-cluster distance, initialization, or hard-assignment rule erases srFGW’s reported gains in normalized total variance (or modularity/silhouette) relative to Louvain and Snake spectral clustering at matched structure–attribute balance, the claimed advantage of joint OT fusion would not hold.","supporting_citations":[],"review_version":1}