{"id":"59755552-1bf6-4d98-9643-a278b310e8f0","arxiv_id":"2412.01122","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"TAS-TsC, a tri-space framework of temporal, attribute, and spatial features, improves truck ETA accuracy on Shenzhen GPS trajectories compared to existing methods.","lead":"This paper proposes TAS-TsC, a deep learning framework that combines Mamba state-space models, hand-crafted trajectory statistics, and graph diffusion to predict truck arrival times from GPS data. It reports lower prediction errors than several baselines on a Shenzhen truck trajectory dataset.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The training objective in §3.5.1 is ill-posed: Eq. (20) uses an undefined index j, and Eq. (21) cannot backpropagate through the non-differentiable k-NN graph, so TAS-TsC cannot be trained as described.","rationale":"The strongest claim in the paper is the empirical result that TAS-TsC outperforms all baselines. My stress-test focuses on the prior condition: the method must be trainable as described. I examined Section 3.5.1 because it defines the only learning signal for the temporal module. Eq. (20) is nonsensical without a specification of j; if j equals i, the term vanishes, and if j is a neighbor, the k-NN selection is non-differentiable. Eq. (21) likewise cannot propagate gradients from LS to the Mamba parameters because ES depends on W, which is built via the argmin operation of Eq. (17); the gradient of argmin with respect to its input is zero almost everywhere. Thus the combined loss LSE in Eq. (22) does not provide a valid training signal for the temporal encoder. A working codebase would have to depart from the paper, meaning the reported results cannot be attributed to the described mechanism. This is an internal inconsistency, not a disagreement with external consensus. The reader's weakest assumption concerned the semantic validity of the k-NN graph; I partially agree because both concerns target the SFM and self-supervised pipeline, but the loss ill-posedness is more fundamental: even with a perfect graph, the model could not be trained. My proposed test is to re-implement the training loop and check the gradient path; if the gradient is zero or the loss is undefined, the central claim is unsupported. The verdict should remain REJECT; no adjustment is needed. I also note that the abstract lists four improvement percentages for three metrics, and the claimed 'second-best baseline IGT' is contradicted by Table 2 (XGB has lower MSE and RMSE on the All dataset), further undermining precision, but these are secondary to the trainability failure.","tokens_in":20853,"tokens_out":10699,"duration_ms":91184,"concrete_test":"Implement the training loop exactly as specified in Section 3.5.1: compute ET from Mamba, build the k-NN graph via Eq. (17), obtain ES via Eq. (19), evaluate LSE from Eqs. (20)-(22), and attempt to backpropagate. If Eq. (20) is undefined or the gradient of LSE with respect to the Mamba parameters is zero because of the argmin in Eq. (17), the method fails as described. Request the authors' code and compare the actual training objective against the paper. Alternatively, independently derive dLS/dtheta_TLM and show it is zero; if so, the claimed joint training is impossible. This test settles whether the reported results stem from the described tri-space coordination or from an unreported modification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is the empirical superiority of TAS-TsC on the Shenzhen truck ETA task. For that claim to hold, the framework must be implementable and trainable exactly as described. Section 3.5.1 defines two self-supervised losses. Eq. (20) contains an undefined index j: if j equals i, the loss is identically zero, and if j denotes a neighbor, the k-NN selection in Eq. (17) is non-differentiable, so no gradient reaches the Mamba parameters through this term. Eq. (21) is also problematic: it computes (ET_i)^T times a weighted sum of Euclidean distances over ES and then multiplies by ET_i, yielding ||ET_i||^2 times a graph distance, but ES is produced by diffusing EA over the adjacency matrix W, and W is constructed by the argmin operation in Eq. (17). That argmin has zero gradient almost everywhere with respect to ET, so LS provides no gradient to the temporal encoder. The paper states in Eq. (22) that LSE = LE + eta*LS is optimized, but no differentiable path exists from LS to the TLM parameters. The only remaining term, LE, does not enforce the stated goal of preserving similarity between ET and XT because the j index is undefined. Thus the described architecture cannot be trained end-to-end. Any working implementation must use a different loss, a differentiable graph construction, or an unreported training schedule. Consequently, the reported 8.7%, 14.1%, 11.2%, and 23.6% improvements cannot be attributed to the framework as written, and the central claim is unsubstantiated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TAS-TsC, a framework for truck arrival-time estimation from GPS trajectories, based on three cooperating feature spaces: a temporal module (TLM) built on the Mamba state-space model, an attribute extraction module (AEM) that reduces variable-length trajectories to fixed statistical features, and a spatial fusion module (SFM) that builds a k-nearest-neighbor graph on temporal embeddings and diffuses attribute features over it. A downstream module (DPM) combines attribute and spatial embeddings with Histogram-based Gradient Boosting to predict arrival times. The framework is evaluated on truck trajectories from five Shenzhen districts, with claims of state-of-the-art performance on MSE, RMSE, MAPE, and MAE, additional cross-domain generalization experiments, and ablations of TLM, AEM, and SFM components.","tokens_in":21261,"tokens_out":7334,"duration_ms":66464,"significance":"If the claims hold, the paper would make a useful applied contribution by demonstrating that state-space sequence models and graph diffusion over trajectory embeddings can improve ETA on sparse, variable-length GPS data. The tri-space decomposition is sensible, and the use of a real truck fleet dataset with five districts and a cross-domain setting is a strength. However, the significance is currently limited by an ill-defined self-supervised training objective, internal inconsistencies between the text and the reported numbers, and the absence of statistical significance testing or repeated-run variability analysis. No code or data availability statement is provided, which hampers reproducibility.","major_comments":[{"comment":"The embedding-learning loss LE contains an undefined index j. As written, the sum over i refers to ET_j and XT_j without specifying how j is selected; if j equals i, the loss is identically zero, and if j is intended to denote a different trajectory or a neighbor, that pairing is not defined. Since LSE = LE + eta*LS (Eq. 22) is the sole self-supervised objective for the temporal encoder before the HGB predictor is fit, Eq. (20) must be corrected and the pipeline retrained before the reported results can be attributed to the described framework.","section":"§3.5.1, Eq. (20)"},{"comment":"The structural loss LS is dimensionally and semantically unclear. (ET_i)^T times a scalar (the sum over j of Euclidean distances between spatial embeddings ES_i and ES_j) times ET_i equals ||ET_i||^2 times that scalar, which is not a natural 'weighted distance based on the difference in attribute embeddings.' Moreover, ES is produced by diffusing EA over the adjacency W, and W is built from a hard argmin-k selection in Eq. (17); the selection operation has no gradient, so the learning signal reaching the temporal encoder through the graph is at best a subgradient through the selected edge weights. Please either relax the graph construction (e.g., soft k-NN), specify a fully differentiable path, or present evidence that LS actually trains TLM as claimed.","section":"§3.5.1, Eq. (21)"},{"comment":"The reported improvements do not match the table. On the 'All' dataset, IGT has MSE 0.0254, RMSE 0.1565, MAPE 0.4966, and MAE 0.1328, while TAS-TsC has MSE 0.0181, RMSE 0.1345, MAPE 0.4410, and MAE 0.1014, giving reductions of approximately 28.7%, 14.1%, 11.2%, and 23.6%, respectively. The text claims '8.7%, 14.1%, 11.2%, and 23.6% improvement ... on MSE, RMSE, and MAPE,' which is internally inconsistent and lists four percentages while naming only three metrics.","section":"§4.2, Table 2"},{"comment":"The empirical central claim rests on a single 7:1:2 split with no repeated runs, error bars, confidence intervals, or significance tests, and the 'strict parameter tuning' of each method is not documented. Without such statistics, the reported improvements over the second-best baseline cannot be distinguished from random variation; please add repeated-seed results and statistical comparisons for the main tables and the cross-domain table.","section":"§4.2 and Table 3"},{"comment":"The spatiotemporal relation graph is constructed by k-NN on flattened temporal embeddings of trajectories padded to Mmax = 1999, while the mean trajectory length is as low as 546.84 (Table 1). The graph may therefore encode trajectory length and padding structure rather than genuine inter-truck interactions such as encounters or shared traffic state. The SFM ablations in Table 5 only show that removing the diffusion or the structural loss hurts performance; they do not validate the semantics of the learned graph. Please provide diagnostics such as edge overlap with spatial proximity, sensitivity to the padding strategy, or a controlled experiment that demonstrates the graph captures meaningful trajectory interdependencies.","section":"§3.4.1 and §4.4"}],"minor_comments":[{"comment":"Section 4.1.3 lists MSE, RMSE, MAPE, and MAE as evaluation metrics, while Section 4.2 says 'Table 2 presents the results across three evaluation metrics (MSE, RMSE, and MAPE)'; the table actually reports four metrics, so the text should be aligned.","section":"§4.1.3 and §4.2"},{"comment":"The sentence introducing the feature diffusion method contains the typo 'we appily adjacency matrix W' instead of 'we apply adjacency matrix W.'","section":"§3.4.2, Eq. (18)"},{"comment":"The keyword list contains 'Tir-space Coordination,' which should be 'Tri-space Coordination.'","section":"Keywords and Abstract"},{"comment":"The complexity of SFM is stated as O(NM^2), but a dense pairwise Euclidean distance computation on N trajectories with M-dimensional features is O(N^2 M); please clarify whether approximate nearest-neighbor methods are used or correct the complexity statement.","section":"§3.6"},{"comment":"The 'residual connection' in EH = EA + alpha*ES is a fixed addition rather than a learned residual block; the terminology may mislead readers.","section":"§3.5.2, Eq. (23)"},{"comment":"Reference [55] is a self-citation appearing in a general list of graph-based models for ETA; it is not clearly connected to the ETA problem and should be replaced with a directly relevant trajectory-graph work or justified in the text.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of an applied soft computing or intelligent transportation journal. The main issue for the editor is that the self-supervised training objective, which is the only training signal for the temporal encoder, is not fully defined as written; this is fixable but requires re-running the experiments after redefining the loss. I also recommend asking the authors to provide repeated-seed statistics and a data/code availability statement, since the empirical claims are central. The self-citation [55] in the related-work section appears tenuously related and could be seen as citation padding; I would ask the authors to justify its relevance or remove it."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is an applied ETA model for truck trajectories, combining Mamba, hand-crafted statistical attributes, and graph diffusion over a k-NN graph. The specific tri-space combination appears new, and the authors tackle a real problem: sparse, variable-length GPS sequences with interdependencies between trucks. The domain-generalization experiment (train on two districts, test on two others) is a good addition, and the ablation study covers the major components. The literature review is current and broad.\n\nThe empirical claim, however, rests on a training objective that is broken as written. In §3.5.1, Eq. (20) uses an undefined index j; if j=i the term is identically zero, and if j denotes a neighbor the k-NN selection in Eq. (17) is non-differentiable. Eq. (21) multiplies (ET_i)^T by a weighted distance sum over ES and then by ET_i, yielding ||ET_i||^2 times a graph distance. The graph W is built by an argmin, so LS gives no gradient to the temporal encoder. The paper states LSE = LE + η LS is optimized, but no differentiable path exists from LS to the TLM parameters. The only well-defined term, LE, doesn't enforce the stated goal because of the undefined j. So the described architecture cannot be trained end-to-end. Any working implementation must use a different loss or a differentiable graph construction.\n\nThere are also smaller reporting problems: the abstract lists four improvement percentages (8.7%, 14.1%, 11.2%, 23.6%) but names three metrics; Table 2 shows the MSE improvement is 28.7%, not 8.7%. No error bars or significance tests are provided, and the dataset and code are not released.\n\nWhat's good: the problem framing is clear, the modules are sensible at a high level, and the idea of modeling multi-truck interaction via graph diffusion is worth exploring. But the central claim—that TAS-TsC beats IGT by these margins—is not supported by the paper's own equations.\n\nI would not send this to referees until the authors fix the training objective, correct the abstract numbers, and release code/data. If those are addressed, the empirical claim could be tested and the paper might become publishable.","headline":"The paper reports large ETA gains from a Mamba+graph-diffusion truck ETA model, but the training objective in §3.5.1 is not differentiable as written, so the central empirical claim is unsubstantiated until the method is fixed.","tokens_in":21719,"tokens_out":5687,"would_cite":false,"duration_ms":44722,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Truck arrival-time forecasts improve when GPS trajectories are embedded jointly as temporal sequences, attribute statistics, and graph-diffused spatial relations, and the paper reports the best results on the tested Shenzhen truck data.","keywords":["Estimating Time of Arrival","GPS Trajectory Data","Tri-space Coordination","Temporal Learning","Attribute Extraction","Spatial Fusion","State Space Model","Graph Representation Learning"],"falsifier":"Rewire the graph's edges randomly while preserving each node's degree and retrain the full framework on the Shenzhen data: if the combined-dataset RMSE stays near 0.1345, the graph diffusion is not the source of the reported gain; a complementary check is whether graph edges connect trajectories whose GPS points actually come within a few hundred meters and a few minutes of each other.","tokens_in":20672,"feed_emoji":"🚚","tokens_out":10197,"duration_ms":86204,"temperature":0.7,"pith_summary":"This paper argues that truck arrival-time estimates become more accurate when GPS trajectories are represented in three complementary spaces instead of one: a temporal space learned by a state-space sequence model, an attribute space of statistical summaries, and a spatial space built from a graph of similar trajectories. The proposed TAS-TsC framework combines these into a low-dimensional hybrid embedding and feeds that embedding to a histogram-based gradient-boosting regressor. The authors test the framework on GPS trajectories from 6,487 truck trips across five Shenzhen districts and report that it achieves the lowest error on every district and on the combined dataset, beating the strongest hybrid baseline by about 14% on RMSE and 11% on MAPE on the combined set. The practical point is that logistics planners could get better arrival-time forecasts from GPS data alone, without map matching or explicit traffic information. The central claim is therefore an architectural one: tri-space coordination, with a Mamba temporal encoder and graph diffusion of attribute features, is what drives the accuracy gain.","feed_headline":"Graph-diffused GPS trajectories sharpen truck ETA","feed_subtitle":"Mamba temporal encoding plus attribute and spatial fusion beats strong ETA baselines on Shenzhen truck data.","key_machinery":"The load-bearing object is the hybrid embedding $E_H = E_A + \\alpha E_S$, formed by a residual connection between the attribute embedding $E_A$ and the diffused spatial embedding $E_S$, which is then passed to a Histogram-based Gradient Boosting predictor. $E_A$ is a 24-dimensional vector of per-trajectory statistics: time differences and rates, longitude and latitude ranges and centers, speed and speed-rate statistics, direction differences and discretized angles, and event differences. $E_S$ comes from the Spatial Fusion Module, which builds a spatiotemporal relation graph $G=\\{W,E_A\\}$ whose adjacency $W$ links each trajectory to its $K$ nearest neighbors under Euclidean distance in the Mamba temporal embedding space, then iteratively propagates features as $E_S^{(l+1)}=D^{-1/2}WD^{-1/2}E_S^{(l)}$. Two self-supervised losses shape the temporal embeddings so that cosine similarity tracks raw-sequence similarity and the structure of spatial differences is reflected in the temporal embedding. The mechanism's job is to let one truck's predicted arrival time absorb information from other trucks that move similarly, while keeping the final regressor's input small and cheap.","core_discovery":"The central claim is that ETA for trucks is best modeled as a tri-space coordination problem rather than as a pure sequence or pure graph problem. The Temporal Learning Module encodes raw GPS sequences with Mamba; the Attribute Extraction Module summarizes each trajectory by max, min, mean, and variance of time differences, speeds, directions, and event flags into a 24-dimensional vector; and the Spatial Fusion Module builds a K-nearest-neighbor graph over the temporal embeddings and propagates attribute vectors through that graph with normalized diffusion. The final prediction is made by Histogram-based Gradient Boosting on the attribute embedding plus a scaled spatial embedding. On the combined Shenzhen dataset, the paper reports the best MSE, RMSE, MAPE, and MAE among all compared methods, with RMSE 0.1345 and MAPE 0.4410 versus 0.1565 and 0.4966 for the second-best baseline IGT.","pith_inferences":["If the graph's neighbor links are truly semantic, the same tri-space recipe should transfer to other GPS-only ETA settings such as buses, ride-hailing, or delivery vans; the paper does not test those settings.","A testable extension the authors leave implicit is to build the graph from actual road-network or encounter proximity instead of Euclidean temporal-embedding distance; the change in error would isolate how much of the gain is genuine spatial interaction rather than embedding similarity.","The self-supervised losses could be reused as a general pretraining objective for trajectory encoders, independent of the final gradient-boosting regressor.","Since the final predictor only sees a small hybrid embedding rather than raw sequences, the practical bottleneck at city scale will be the K-nearest-neighbor graph construction, which has $O(lNM^2)$ complexity in the paper's analysis; approximate neighbor search would be the natural scaling fix."],"forward_implications":["ETA can be improved using only raw GPS points (time, coordinates, speed, direction, event flag), with no map-matched road segments or external traffic feeds.","The Mamba-based temporal encoder provides linear-time sequence modeling, so the framework scales better to long trajectories than Transformer-based ETA models.","Because the spatial module propagates information over a graph of similar trajectories, the predicted arrival time of a truck implicitly reflects conditions of neighboring trips, which is where congestion and merging effects show up.","The cross-district experiments indicate that embeddings learned in one set of districts transfer to unseen districts, so the method could be deployed in cities where no local training labels exist."],"supporting_citations":[{"why":"Supplies the Mamba selective state space model used by the Temporal Learning Module to encode sparse, variable-length GPS sequences.","marker":"[43]"},{"why":"Supplies the Histogram-based Gradient Boosting algorithm used by the Downstream Prediction Module to map the hybrid embedding to arrival time.","marker":"[45]"},{"why":"Defines the Inductive Graph Transformer baseline that the paper reports as its strongest competitor on the combined dataset.","marker":"[10]"},{"why":"Supplies the MetaTTE hybrid baseline used in the main comparison and domain-transfer experiments.","marker":"[16]"},{"why":"Provides the normalized graph convolution formulation $D^{-1/2}WD^{-1/2}$ that the Spatial Fusion Module adapts for feature diffusion.","marker":"[40]"},{"why":"Supplies the residual connection idea used to fuse attribute and spatial embeddings into the hybrid embedding.","marker":"[44]"}],"fun_headline_variants":["Mamba plus graph diffusion fuse truck GPS for sharper ETA","Tri-space coordination cuts truck ETA error on Shenzhen data","TAS-TsC: time, attributes, space unify for better truck ETA","Graph-propagated attributes boost truck arrival time forecasts","Truck ETA improved via temporal, attribute, spatial tri-space fusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the K-nearest-neighbor graph built from Euclidean distances between temporal embeddings corresponds to real interdependencies among trucks (encounters, merges, shared traffic conditions); if that graph mostly reflects padding structure or noise, the spatial fusion module's contribution collapses.","fun_headline_variants_meta":{"raw":{"variants":["Mamba plus graph diffusion fuse truck GPS for sharper ETA","Tri-space coordination cuts truck ETA error on Shenzhen data","TAS-TsC: time, attributes, space unify for better truck ETA","Graph-propagated attributes boost truck arrival time forecasts","Truck ETA improved via temporal, attribute, spatial tri-space fusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000264,"raw_usage":{"total_tokens":1595,"prompt_tokens":925,"completion_tokens":670,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":578}},"tokens_in":541,"tokens_out":670,"duration_ms":6873,"temperature":1.0,"reasoning_tokens":578,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:39:42.591712+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rewire the graph's edges randomly while preserving each node's degree and retrain the full framework on the Shenzhen data: if the combined-dataset RMSE stays near 0.1345, the graph diffusion is not the source of the reported gain; a complementary check is whether graph edges connect trajectories whose GPS points actually come within a few hundred meters and a few minutes of each other.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Histogram-based Gradient Boosting algorithm used by the Downstream Prediction Module to map the hybrid embedding to arrival time."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Inductive Graph Transformer baseline that the paper reports as its strongest competitor on the combined dataset."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the MetaTTE hybrid baseline used in the main comparison and domain-transfer experiments."}],"review_version":1}