{"id":"dd52145c-489e-4ab6-bc3b-b1695d098b4d","arxiv_id":"2411.14768","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Jointly encoding grid and road trajectory expressions with cross-attention fusion outperforms single-view trajectory representation learning on three downstream tasks.","lead":"GREEN is a trajectory representation learning method that fuses grid and road views of GPS trips using contrastive and masked-language modeling losses. On Porto and Chengdu taxi data it reports consistent gains over seven baselines for travel time estimation, trajectory classification, and similarity search, although the results lack error bars.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The largest MSTS gains (46.99% and 55.68% MR reductions) are measured against a self-constructed benchmark whose 'positive' is the same trajectory after downsampling; this tests robustness to GPS downsampling, not general trajectory similarity.","rationale":"The reader's weakest-assumption analysis identifies the same vulnerability: the MSTS benchmark is constructed by the authors and validated against no external similarity measure. My reading of Appendix A.3 confirms that the positive candidate is an augmented version of the query trajectory itself, filtered by a change-rate heuristic. The very large MR reductions reported for GREEN are therefore plausibly explained by the model's robustness to the specific downsampling transformation, not by superior general similarity discrimination. The concern is load-bearing because MSTS contributes the largest component of the advertised average improvement; without it, the headline 15.99% drops to roughly 11-12%, though the TTE and TC improvements still support a conditional acceptance. I do not see an internal inconsistency that would justify rejection. The right response is to keep the verdict CONDITIONAL and require a check of MSTS against an independent similarity ground truth or, at minimum, a sensitivity analysis over the downsampling ratio and change-rate filter.","tokens_in":21915,"tokens_out":6026,"duration_ms":65393,"concrete_test":"Replace the Appendix A.3 heuristic with an externally grounded similarity benchmark: for each query trajectory, label as positive any database trajectory whose road-segment Jaccard overlap or full-GPS Hausdorff/DTW distance exceeds a pre-specified threshold, and additionally test on a small set of human- or domain-expert-validated similar pairs if available. Recompute MR and HR@1 for GREEN and JGRM on this benchmark. If GREEN's MR advantage falls below roughly 10% relative or reverses, the MSTS entries in Tables 2-3 should not be used to support the generality claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim of an average 15.99% improvement is dominated by the most-similar-trajectory-search results in Tables 2 and 3, where GREEN reduces mean rank by 46.99% (Porto) and 55.68% (Chengdu) over the best baseline. The ground truth for this task, described in Appendix A.3, is not an independently validated similarity measure. For each query, the positive candidate is generated from the same raw GPS trajectory by keeping start and end points fixed, downsampling the remaining GPS points by 0.5, re-map-matching, and then filtering to road trajectories whose change rate lies between 0.3 and 0.5. The task therefore measures whether a representation is robust to this specific downsampling-and-map-matching perturbation, not whether it returns trajectories that are semantically similar in any general sense. Because GREEN is explicitly trained to align grid and road expressions of the same raw trajectory, it is unsurprising that it wins a benchmark where the positive is a perturbed version of the query. The reported MR improvements cannot be interpreted as evidence that GREEN finds more similar trajectories in the usual retrieval sense. Recomputing the headline average without the two MR entries lowers it to roughly 11-12%, so the TTE and TC improvements remain, but the largest claimed advantage is unsubstantiated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GREEN, a self-supervised trajectory representation learning method that jointly uses grid and road trajectory expressions. It builds a CNN-plus-Transformer grid encoder and a GNN-plus-Transformer road encoder, trains them with a contrastive loss and a masked-language-model-style loss in which grid trajectories assist the reconstruction of masked road segments, and fuses the two views with a cross-attention interactor. The evaluation on Porto and Chengdu compares GREEN with seven baselines on travel time estimation, trajectory classification, and most similar trajectory search, and reports consistent improvements, together with ablations, a pre-training data-efficiency study, an efficiency comparison, and a transferability analysis.","tokens_in":22213,"tokens_out":6908,"duration_ms":69336,"significance":"The core observation that grid and road expressions carry complementary information is well motivated, and the model is described in sufficient detail for reproduction. The self-supervised losses are not fitted to downstream labels, and the pre-training data-efficiency experiment in Figure 5 is a useful sanity check. If the empirical claims are validated, the work would be a solid contribution to trajectory representation learning, and the public code and data release are a clear strength. However, the evaluation currently contains a load-bearing caveat around the most-similar-trajectory-search benchmark and lacks statistical grounding, so the magnitude and generality of the claimed improvements are not yet firmly established.","major_comments":[{"comment":"The 'most similar trajectory search' benchmark is constructed by taking each query's own raw GPS trajectory, keeping the endpoints, downsampling the remaining points by 0.5, re-map-matching, and retaining candidates whose road change rate is between 0.3 and 0.5. As constructed, this task measures robustness to a specific downsampling-and-map-matching perturbation rather than general trajectory similarity, and the ground truth is not validated against any external similarity measure. Because GREEN is explicitly trained to align the grid and road expressions of the same raw trajectory, it has a structural advantage on exactly this kind of perturbation. The MR reductions of 46.99% and 55.68% therefore cannot be interpreted as evidence that GREEN retrieves more semantically similar trajectories in general. Moreover, the abstract's headline average improvement of 15.99% is dominated by these two MR entries; excluding them leaves roughly 11-12%. Please either reposition the MSTS claim as a downsampled-trajectory retrieval result or validate the benchmark against independent ground truth, for example human judgments or a standard trajectory similarity measure such as DTW/LCSS.","section":"Appendix A.3; Tables 2-3"},{"comment":"All accuracy numbers are single runs with no standard deviations, confidence intervals, or significance tests, and the hyperparameter study in Appendix A.4 appears to select values (grid cell size, mask ratio and length, dimensions, layer counts) using the same evaluation metrics reported in the main tables, without any statement that a separate validation split was used. With improvements as small as 1.26% and 1.79% on Chengdu trajectory classification, it is difficult to tell whether the reported margins are genuine or are artifacts of run-to-run variance and test-set selection. Please report means and standard deviations over multiple seeds, and state explicitly whether any hyperparameter was chosen on the test set.","section":"Section 5.1 and Appendix A.4; Tables 2-4"},{"comment":"The transferability section claims that the transferred model 'still outperforms most baselines,' but Table 5 only compares the w/ Transfer and w/o Transfer variants and contains no baseline rows. Either add the corresponding baseline results to the table or soften the claim, because the current statement is not supported by the reported evidence.","section":"Section 5.3, Table 5"},{"comment":"The 'w/o MLM loss' variant removes both the MLM loss and the dual-modal interactor, replacing the learned fusion with a simple average of the two encoder outputs. The observed degradation therefore conflates the effect of the MLM objective with the effect of the cross-attention fusion module, so the text's conclusion that 'the MLM loss is more important for trajectory classification than the contrastive loss' is not supported by this ablation. A cleaner ablation would remove only the MLM loss while retaining the interactor, or otherwise separate the two design choices.","section":"Section 5.3, Table 4"}],"minor_comments":[{"comment":"Equation (14) appears to compute the cross-entropy over all positions in the road trajectory, while the text states that only masked road segments are predicted. Please clarify whether unmasked positions contribute to the loss and how masked positions are represented in the output.","section":"Section 4.4, Eq. (14)"},{"comment":"The mask-recovery curves in Figure 3 are used to justify the asymmetric MLM design, but the experimental setup for this figure is not described anywhere (dataset split, number of runs, definition of accuracy on masked tokens). Please add this information so the figure can be interpreted.","section":"Section 4.4, Figure 3"},{"comment":"The statement that 'less than 1% of the grids under 100m×100m in the dataset have more than one GPS point' is presented without a derivation or a specified dataset; please provide the computation or a citation backing this statistic.","section":"Section 4.1"},{"comment":"The term 'accuracy' in the abstract is used as an umbrella for MAE/MAPE/RMSE, F1 scores, and retrieval metrics; consider replacing it with a more neutral phrase such as 'performance' to avoid implying a single accuracy measure.","section":"Section 5.2, Tables 2-3"}],"recommendation":"major_revision","confidential_remarks":"The main gate for acceptance is the MSTS benchmark: the authors should either validate it externally or reframe the claim as robustness to downsampling-and-map-matching, and they should add multiple-run statistics. If those points are addressed, the paper is likely to be a solid empirical contribution. I would not require additional theory; the model and ablations are otherwise reasonable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: GREEN is the first TRL method I know of that feeds both grid and road trajectories into one model, and the travel-time and classification gains look real. The headline \"15.99% average improvement,\" though, is mostly carried by the most-similar-trajectory-search numbers, and those are measured against a benchmark the authors built themselves that basically tests whether a method can recognize a downsampled version of the query. That doesn't sink the paper, but it means you should treat the MR reductions with skepticism.\n\nWhat's genuinely good: the motivation is clear, the architecture is sensible (CNN for grid, GNN+Transformer for road, contrastive loss plus a grid-assisted masked-road reconstruction loss, cross-attention fusion), and the ablations show each piece helps. The MLM design — using grid trajectories to help reconstruct masked road trajectories, rather than the other way around — is a nice touch with a plausible rationale. They also report efficiency and transferability, which is more than most TRL papers do. The code is promised on GitHub.\n\nThe soft spots are in the evaluation. No error bars or multiple runs anywhere, so we have no sense of variance. Hyperparameter choices (grid size, mask ratio, embedding dimension) appear to be selected on the test set, per Appendix A.4. And the MSTS ground truth in Appendix A.3 is constructed by downsampling the query trajectory, re-map-matching, and keeping only cases with a change rate between 0.3 and 0.5. That means the \"positive\" is a perturbed copy of the query, not an independently judged similar trajectory. GREEN is explicitly trained to align grid and road views of the same raw trajectory, so it's not surprising it wins at recognizing a downsampled version. If you remove the two MR entries, the average improvement drops to roughly 11-12%, which is still respectable. The TTE gains (12-27% MAE reduction) and TC gains (1.3-3.1% F1) are the more trustworthy part.\n\nWho should read this: anyone working in trajectory representation learning or spatial-temporal self-supervision. It's a solid empirical contribution with a novel combination, but the evaluation needs tightening before I'd trust the headline number. A serious referee should push for error bars, a validated similarity benchmark, and clarification of how hyperparameters were chosen. I'd accept it for review — it's not a desk reject — but expect revisions.","headline":"A genuinely novel combination of grid and road trajectory views with solid TTE/TC gains, but the headline similarity-search numbers rest on a self-built benchmark that overstates the method.","tokens_in":22743,"tokens_out":2765,"would_cite":true,"duration_ms":26150,"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":"GREEN claims that jointly encoding grid and road trajectory views yields trajectory representations that consistently beat single-view methods on three downstream tasks.","keywords":["Trajectory representation learning","Grid trajectory","Road trajectory","Multimodal learning","Contrastive learning","Masked reconstruction","Travel time estimation","Trajectory similarity"],"falsifier":"Re-run the most similar trajectory search experiments with an independent ground truth, for example human-annotated similar trajectory pairs or a classic GPS-level similarity measure such as dynamic time warping, and check whether GREEN's mean-rank reduction over JGRM persists; if the ranking advantage shrinks or reverses, the key claim about similarity search would not survive.","tokens_in":21729,"feed_emoji":"🚕","tokens_out":5756,"duration_ms":49982,"temperature":0.7,"pith_summary":"GREEN is built on the observation that a raw GPS trajectory can be expressed in two complementary forms: a grid trajectory, which records the regions and locations it passes through in free space, and a road trajectory, which records the sequence of road segments in the underlying network. The paper argues that existing trajectory representation learning (TRL) methods exploit only one of these forms and therefore throw away information the other form provides. Its central claim is that jointly encoding both expressions, aligning them with a contrastive loss, and using the grid expression to help reconstruct masked road segments yields trajectory vectors that are more accurate for downstream tasks than vectors learned from either expression alone. Across travel time estimation, trajectory classification, and most similar trajectory search on the Porto and Chengdu datasets, the method reports consistent gains and an average 15.99 percent improvement over the best baseline.","feed_headline":"Trajectory learning jumps 16% when grid and road views merge","feed_subtitle":"GREEN fuses free-space cells and road-network segments, improving travel time, classification, and similarity search.","key_machinery":"The central machinery is the two complementary trajectory encoders plus the two alignment losses. Each raw GPS trajectory is converted into a grid trajectory and a road trajectory; the grid encoder uses a CNN on a three-channel grid map (coordinates and traffic flow) followed by a Transformer, and the road encoder uses a graph attention network on road-segment attributes followed by a Transformer whose attention is adjusted by road-type continuity. A contrastive loss encourages the grid and road embeddings of the same trajectory to land near each other, and a masked-language-model loss makes the grid representation help the model recover segments masked from the road representation; this second loss is what forces the grid view to supply information the road view lacks. The dual-modal interactor, a cross-attention layer with queries from the road view and keys and values from the grid view, fuses both views into the final vector.","core_discovery":"The paper's discovery is that the two trajectory expressions are not redundant but complementary, and that this complementarity can be exploited by a dual-encoder architecture. The grid encoder applies a CNN over cell traffic-flow and coordinate channels, then feeds a Transformer with trajectory-specific GPS and time features; the road encoder applies a graph attention network and a Transformer with coarse-grained day-of-week and minute-of-day encodings, and injects road-type continuity into the attention matrix. The two encoders are aligned by a contrastive loss and by a masked-language-model loss in which grid representations assist in reconstructing masked stretches of road trajectories. A dual-modal interactor then fuses the two views through cross-attention to produce the final trajectory representation. The author would state the result as: for travel time estimation, trajectory classification, and most similar trajectory search, GREEN consistently outperforms seven strong TRL baselines on two real-world datasets, with larger gains on tasks that depend directly on the learned representations.","pith_inferences":["The grid-helps-road masking idea is a template for any pair of trajectory expressions with unequal reconstruction difficulty: the easier-to-reconstruct view (here, road segments with network continuity) should be the reconstruction target, and the harder view should be the prompt; the paper argues the opposite direction fails because discrete grids are hard to recover.","If the most-similar-trajectory ground truth (downsampling by 0.5 and keeping road trajectories with change rate 0.3-0.5) is replaced by an established similarity measure, GREEN's large reported improvement in similarity search would need to be rechecked, because that improvement may partly reward robustness to downsampling rather than general similarity accuracy.","The two-view pretraining objective could transfer to other mobility representations, such as semantic maps or point-of-interest sequences, whenever two views of the same movement carry complementary information; this extension is not tested in the paper.","Because the ablations show grid-only and road-only variants trade off differently across tasks (grid helps classification and travel time, road helps similarity search), a practitioner could tune the fusion weight per task rather than using fixed cross-attention fusion."],"forward_implications":["Travel time estimation improves by 12.26% to 26.79% in MAE over the best baseline on the two datasets, because the fused representation carries both region-level and road-level timing information.","Trajectory classification improves by smaller margins (1.26% to 3.11%), which the paper attributes to classification being an easier task, but GREEN still leads across both datasets.","Most similar trajectory search shows the largest relative gains on the author-built benchmark: mean rank drops by 46.99% on Porto and 55.68% on Chengdu compared with the best baseline.","Pre-training with the two losses matches the accuracy of training from scratch on the full Chengdu dataset using only 30% of the training data.","Training time is shorter than the strongest road-based baselines and inference time is comparable, so the accuracy gain does not come with an obvious computational penalty."],"supporting_citations":[{"why":"Supplies the state-of-the-art grid-based contrastive baseline (TrajCL) that GREEN is compared against.","marker":"[4]"},{"why":"Provides the masked language model pre-training technique that GREEN adapts into its masked-road reconstruction loss.","marker":"[10]"},{"why":"Supplies the road-based RNN seq2seq baseline (Trembr) used as a comparison.","marker":"[13]"},{"why":"Supplies the BERT-style road trajectory baseline (START) that is among the strongest comparisons.","marker":"[17]"},{"why":"Supplies the grid trajectory representation method (t2vec) that motivates grid-based encoding.","marker":"[24]"},{"why":"Supplies the joint GPS-route road baseline (JGRM) that is the strongest comparison for similarity search.","marker":"[30]"},{"why":"Supplies the road-network contrastive baseline (JCLRNT) used in the comparison.","marker":"[31]"},{"why":"Supplies the Transformer backbone used by both encoders and the interactor.","marker":"[34]"},{"why":"Supplies the graph attention network that learns road segment embeddings from the road network.","marker":"[35]"},{"why":"Supplies the map-matching algorithm that converts raw GPS trajectories into road trajectories.","marker":"[38]"}],"fun_headline_variants":["Grid and road views together lift trajectory learning by 16%","Merging grid and road trajectories boosts representation learning","GREEN fuses grid and road data to improve trajectory tasks","Combining two trajectory views yields 16% accuracy gain","Dual-view trajectory learning beats state-of-the-art by 16%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the author-built ground truth for most similar trajectory search, which downsamples GPS points by 0.5, map matches them again, and keeps only road trajectories with a change rate between 0.3 and 0.5 as positives, measures real trajectory similarity; if this heuristic favors methods that tolerate downsampling, the large reported similarity-search improvements would not reflect general accuracy.","fun_headline_variants_meta":{"raw":{"variants":["Grid and road views together lift trajectory learning by 16%","Merging grid and road trajectories boosts representation learning","GREEN fuses grid and road data to improve trajectory tasks","Combining two trajectory views yields 16% accuracy gain","Dual-view trajectory learning beats state-of-the-art by 16%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000405,"raw_usage":{"total_tokens":2117,"prompt_tokens":968,"completion_tokens":1149,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":1064}},"tokens_in":584,"tokens_out":1149,"duration_ms":10082,"temperature":1.0,"reasoning_tokens":1064,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:54:53.320871+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the most similar trajectory search experiments with an independent ground truth, for example human-annotated similar trajectory pairs or a classic GPS-level similarity measure such as dynamic time warping, and check whether GREEN's mean-rank reduction over JGRM persists; if the ranking advantage shrinks or reverses, the key claim about similarity search would not survive.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the state-of-the-art grid-based contrastive baseline (TrajCL) that GREEN is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the road-based RNN seq2seq baseline (Trembr) used as a comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the BERT-style road trajectory baseline (START) that is among the strongest comparisons."},{"cited_title":"Jensen, and Wei Wei","cited_arxiv_id":null,"evidence_quote":"Supplies the grid trajectory representation method (t2vec) that motivates grid-based encoding."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the joint GPS-route road baseline (JGRM) that is the strongest comparison for similarity search."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the road-network contrastive baseline (JCLRNT) used in the comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the map-matching algorithm that converts raw GPS trajectories into road trajectories."}],"review_version":1}