{"id":"d7efebbd-425c-441b-bae8-b1709116c71f","arxiv_id":"2412.19092","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A graph-enhanced sequential model that fuses long-, mid-, and short-term user preferences reports the best next-location prediction accuracy on NYC, Tokyo, and Dallas check-in benchmarks.","lead":"TrajGEOS is a deep learning model that predicts a person's next check-in location by combining a graph of all users' location transitions with week-level reading of that person's own recent history. The authors report top-1 accuracy gains on three public check-in datasets, making the method a candidate new state of the art for next-location prediction.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim is vulnerable because baselines run on differently preprocessed datasets and several hyperparameters are chosen on the NYC test set, with no error bars; the small R@1 margins may not survive an apples-to-apples comparison.","rationale":"I agree with the reader's weakest_assumption. The load-bearing condition for the central claim is that the reported comparisons isolate model quality. That condition is not met: Table I documents non-identical evaluation sets, and the main result is a point estimate with no variance. The magnitudes matter: Dallas's 0.13 pp gap could easily be within seed noise, and NYC's 0.86 pp gap is reported after test-set-driven hyperparameter selection. The architecture has a plausible mechanism and the ablations are internally coherent, but the headline claim of state-of-the-art performance is established only if the comparison is re-run under a common protocol. Since the concern is about evidence quality rather than a logical contradiction in the method, the appropriate outcome is the reader's CONDITIONAL verdict, unchanged. No code is currently available, which makes the required verification harder but does not by itself falsify the model.","tokens_in":18935,"tokens_out":4689,"duration_ms":47359,"concrete_test":"Run TrajGEOS and every baseline on exactly the same train/test split (e.g., the TrajGEOS split) with identical filtered user/location sets and candidate locations, and choose alpha, recent-window length, and GNN depths on a held-out validation split; report mean plus/minus standard deviation over five seeds. If the R@1 advantage over the best baseline drops below one standard deviation or reverses on any of NYC/TKY/Dallas, the SOTA claim should be downgraded.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section V-A asserts that the per-model datasets in Table I \"exhibit minimal differences,\" but the table contradicts this: on Dallas, TrajGEOS uses 1,357 users and 118,069 records while LSTPM uses 954 users and 103,664 records and DeepMove uses 1,193 users and 93,911 records; on NYC, record counts range from 111,968 (DeepMove) to 138,229 (GeoSAN), with different user sets. Since each model is evaluated on its own preprocessed split of chronological weekly sub-trajectories, the actual test instances differ across models, so a model can appear better without being better. This is not a minor point: the reported R@1 advantages over the best baseline are 0.86 pp on NYC, 1.02 pp on TKY, and only 0.13 pp on Dallas, and all results are five-run means with no standard deviations or significance tests. The concern is compounded by Appendix C: alpha (Table IV), recent-trajectory length (Table V), EGraphSAGE depth (Table VII), and GraphSAGE depth (Table VIII) are all chosen by Recall@1 on the NYC dataset, apparently on the test partition, so the reported NYC margin can inherit the best of many configurations. The architecture may still be sound, but the evidence as presented does not yet establish that TrajGEOS outperforms prior work.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes TrajGEOS, a deep model for next-location prediction from LBSN check-in data. The architecture combines three components: (i) a hierarchical graph module that runs an edge-aware graph convolution (EGraphSAGE) on a global trajectory graph built from users' training trajectories, with node features from location ID, category, and coordinates and edge features from transition counts, distance, and 24-hour flow, followed by a GraphSAGE on per-user subgraphs whose visit-frequency-weighted readout is treated as the user's long-term preference; (ii) a sequential module that encodes the current weekly sub-trajectory with a shared GRU to produce a short-term preference; and (iii) an orientation module that adds position encodings to the preceding two-to-three weeks of records and attends over them, using the GRU output as query, to produce a mid-term preference. A multi-task head predicts the next location and the next category with loss L = alpha*Ll + (1-alpha)*Lc. Experiments on NYC, TKY, and Dallas (Gowalla) report Recall@1/5/10 and MRR@10 against FPMC-D/W, DeepMove, LSTPM, GeoSAN, GETNext, and MTNet, claim to outperform all baselines on all three datasets, and include ablations, visualizations, and sensitivity analyses.","tokens_in":19227,"tokens_out":13515,"duration_ms":105543,"significance":"If the performance claim held, TrajGEOS would be a solid incremental contribution to next-location prediction: it packages graph-derived location embeddings with informative edge attributes (distance, transition counts, 24-hour flow), a user-subgraph readout for long-term preference, and a lightweight attention-based mid-term module. The paper deserves credit for constructing the trajectory graph from training records only, for using a user-level chronological split, for reporting five-run averages of its own model, and for publishing sensitivity tables for its main hyperparameters; the ablation study supports the value of the graph module (e.g., TKY R@1 drops from 0.2490 to 0.2017 without it) and of the orientation module, and the implementation is promised for release.","major_comments":[{"comment":"The assertion in Section V-A that the per-model datasets 'exhibit minimal differences, allowing for a meaningful comparison' is contradicted by Table I. On Dallas, TrajGEOS uses 1,357 users / 118,069 records / 5,428 locations, while LSTPM uses 954 / 103,664 / 5,366, GeoSAN uses 2,300 / 142,980 / 5,357, and DeepMove uses 1,193 / 93,911 / 5,346. On NYC, record counts range from 111,968 (DeepMove) to 138,229 (GeoSAN), with TrajGEOS at 131,874, and the user sets differ as well (e.g., LSTPM 1,019 users). Because each model's trajectories are segmented into weekly sub-trajectories and evaluated on the chronological last 20% per user, differing record subsets imply differing test instances, and differing location sets imply differing candidate sets. The reported advantages over the best baseline (0.86 pp on NYC, 1.02 pp on TKY, 0.13 pp on Dallas) are of the same order as these dataset discrepancies. To support the claim that TrajGEOS outperforms the baselines, the authors should run all models under a single shared preprocessing pipeline (including the same user set, record set, and candidate location set for the test portion) or demonstrate on a common subset of test instances that the ranking is unchanged.","section":"Section V-A, Table I"},{"comment":"The hyperparameters alpha, the recent-trajectory length kappa, the number of EGraphSAGE layers, and the number of GraphSAGE layers are each selected by Recall@1 computed on the NYC dataset, and the selected configurations (alpha = 0.7, kappa = 2, depth 2 in both graph modules) reproduce exactly the NYC R@1 = 0.2721 reported in Table II. This means the reported NYC result, and by transfer the TKY and Dallas results, inherit the best of several configurations evaluated on the test partition. The sensitivity spreads are bounded (at most about 1.0 pp for alpha in Table IV and 0.5 pp for kappa in Table V), so the tuning advantage alone is not huge, but it is comparable to the headline margin over the best baseline on NYC (0.86 pp), and the protocol is still methodologically incorrect. In addition, Section III presents kappa = 2 as a fixed definition ('we set kappa = 2'), whereas Table V shows it is an empirically chosen value. The authors should select all hyperparameters on a validation split (e.g., the sub-trajectories immediately preceding the last 20% per user) and report test-set results for the single chosen configuration.","section":"Appendix C, Tables IV, V, VII, VIII"},{"comment":"The paper reports that TrajGEOS results are averages over five independent runs, but no standard deviations are given, and it is not stated whether the baselines were run multiple times or with what variance. Given that the R@1 margins over the best baselines are 0.86 pp (NYC), 1.02 pp (TKY), and 0.13 pp (Dallas), the state-of-the-art claim requires at least a standard deviation or a significance test on identical test instances. This is particularly important because, on NYC, the ablation TrajGEOS-woGraph (R@1 = 0.2622) falls below MTNet (0.2635), so the entire claimed advantage on NYC rests on the graph module whose comparison is affected by the preprocessing differences raised in the first comment.","section":"Section V-B, Table II"}],"minor_comments":[{"comment":"The edge set is written 'E = {ei->j, ei->j, ...}' with the same ordered pair repeated; this should be a list of distinct transitions.","section":"Section III, Definition 3.3"},{"comment":"In the softmax denominator, 'Pk l=1 exp(beta_j)' should be 'Pk l=1 exp(beta_l)', since the summation index is l.","section":"Section IV-D-1, Eq. (13)"},{"comment":"The phrase 'the k-th dim in the f lowof ei->j' has a spacing error, and 'an experiential value used in dropout layers' should read 'an empirical value'.","section":"Section IV-B"},{"comment":"The subscript in 'h^{k-1}_{i-1,s}' appears to be a typo for 'h^{k-1}_{i,s}', given that the preceding term is 'h^{k}_{i,s}'.","section":"Section IV-B, Eq. (8)"},{"comment":"The TKY column header 'Recall10' is inconsistent with 'R@10' used for the other columns in the same table.","section":"Table II"},{"comment":"The text says 'we present our findings in Table A', but the referenced table is labeled Table VIII.","section":"Appendix C-F"},{"comment":"The meaning of the 'Processed' row is unclear, since no model, including TrajGEOS, uses exactly those counts (e.g., NYC 'Processed' has 1,083 users and 139,183 records, while TrajGEOS reports 1,065 users and 131,874 records).","section":"Table I"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope well (human mobility, deep learning, graph neural networks), and the related-work coverage is adequate. The main gap is methodological: the state-of-the-art claim rests on comparing models on non-identical test sets, with hyperparameters selected on the NYC test set and no error bars. The stress-test concern raised in review does land on reading the paper, and Table I plus Appendix C are the concrete evidence. That said, the sensitivity tables show the model itself is fairly stable, and the ablations do support the value of the graph and orientation modules, so I see a plausible path to acceptance if the evaluation is redone with a shared preprocessing pipeline and validation-based tuning. The authors promise to release the implementation; complete code would also help verify the comparability claim. I would not reject the paper, but I would insist on the evaluation fix before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: the method is a reasonable combination of known building blocks, and the paper mostly knows what it is doing. The thing to worry about is the headline claim. “For all datasets, our model outperforms the baseline models” rests on comparisons where each model was evaluated on its own preprocessed dataset, and several hyperparameters appear to be chosen on the NYC test metric. The reported R@1 margins are 0.86 pp on NYC, 1.02 pp on TKY, and 0.13 pp on Dallas.\n\nWhat is actually new: the hierarchical graph convolution (EGraphSAGE on the global trajectory graph, then GraphSAGE on per-user subgraphs) feeding an attention-based orientation module over the past 2–3 weeks, with next-category prediction as an auxiliary task. I don’t see that exact combination in the cited literature. The ablations are decent and honestly reported: removing the graph costs 1.0 pp on NYC and 2.6 pp on Dallas; removing the orientation module costs 0.8 pp on NYC but only about 0.1–0.2 pp on TKY and Dallas. That makes the orientation module the weakest component, and the paper does not hide it.\n\nThe soft spot is the evaluation protocol. Section V-A says the per-model datasets “exhibit minimal differences,” but Table I contradicts that: on Dallas, TrajGEOS uses 1,357 users and 118,069 records while DeepMove uses 1,193 users and 93,911 records and LSTPM uses 954 users; on NYC, record counts range from 111,968 to 139,183. If the test trajectories are not identical across models, the ranking is not apples-to-apples. This is not a minor point when the margin is 0.13 pp. Appendix C compounds it: alpha, recent-window length, EGraphSAGE depth, and GraphSAGE depth are all selected by R@1 on the NYC dataset, with no indication of a validation split, so the reported NYC number inherits the best of many configurations. Five-run means with no standard deviations or significance tests make it hard to know which differences are real.\n\nThere is no code in the v1, which limits reproducibility. The user-level split is reasonable, and the graph construction is sensible. This paper deserves a serious referee because the architecture is plausible and the evaluation can be fixed, but the SOTA claim should not be accepted as stated. A revision with identical test splits across all models, validation-based hyperparameter selection, and error bars would be the right path.","headline":"TrajGEOS is a competent architecture with honest ablations, but the state-of-the-art claim is not established because baselines run on non-identical data and hyperparameters are tuned on the NYC test set.","tokens_in":19758,"tokens_out":4000,"would_cite":true,"duration_ms":37581,"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":"This paper claims that TrajGEOS, a trajectory graph enhanced orientation-based sequential network, outperforms existing state-of-the-art baselines on next location prediction across three real-world LBSN datasets.","keywords":["next location prediction","human mobility","trajectory graph","graph neural networks","attention mechanism","multi-task learning","location-based social networks","sequential recommendation"],"falsifier":"Re-run every baseline and TrajGEOS on one identical user-level 80/20 split of the same filtered check-in data for NYC and Dallas, with identical test weeks and filtering rules, and check whether TrajGEOS still leads on Recall@1 and MRR@10 across five seeds; if the margin collapses, the reported outperformance is not established.","tokens_in":18724,"feed_emoji":"📍","tokens_out":8604,"duration_ms":71305,"temperature":0.7,"pith_summary":"The paper sets out to improve next-location prediction, the task of guessing which place a person will check in next, by exploiting structure hidden in everyone's movement histories. It argues that standard sequential models overlook the relationships between locations and the pull of a user's most recent weeks on their next move. TrajGEOS builds a global trajectory graph from all users' check-in sequences, learns location and user embeddings through hierarchical graph convolution, and adds an orientation module that attends over the past two to three weeks of a user's trajectory to form a mid-term preference. On three public location-based social network datasets, the model reports higher Recall@1 than the strongest baselines, and ablations credit the gain mainly to the graph embeddings and the mid-term preference signal.","feed_headline":"TrajGEOS tops next-location prediction on three city datasets","feed_subtitle":"Adds location-relation and mid-term preference signals to sequential models, lifting Recall@1 on NYC, Tokyo, and Dallas.","key_machinery":"The central object is the trajectory graph, a directed graph whose nodes are locations and whose edges are observed transitions from all users' training histories, each edge annotated with distance, total transition count, and a 24-dimensional hourly flow vector. Hierarchical graph convolution applies an edge-aware message-passing layer on this global graph to produce location embeddings, then applies a second convolution on each user's subgraph and performs a visit-frequency-weighted readout to obtain a long-term user embedding. The orientation module then adds sinusoidal position encodings to the encoded recent weekly sub-trajectories and uses attention, keyed by the short-term recurrent output, to form a mid-term preference; the three preferences and the user embedding are concatenated and fed to two predictors for next location and next category.","core_discovery":"On its own terms, the paper's central claim is that TrajGEOS achieves state-of-the-art next-location prediction on the NYC, TKY, and Dallas check-in datasets, with Recall@1 of 27.2% versus 26.4% for the best baseline on NYC, 24.9% versus 23.8% on Tokyo, and 13.2% versus 13.1% on Dallas, and consistent leads on Recall@5, Recall@10, and MRR@10. The claimed mechanism is a fusion of three preference scales: long-term preference read out from a user's trajectory subgraph with visit-frequency weights, short-term preference from the current week's recurrent encoding, and a mid-term preference produced by an attention-based orientation module over the preceding two to three weeks. These are combined with location embeddings learned from a global trajectory graph whose edges carry distance, transition counts, and 24-hour flow patterns, and with a next-category prediction auxiliary task that shares the same recurrent outputs. Ablation experiments show that removing the graph module or the orientation module lowers accuracy, while removing the short-term preference causes the largest drop.","pith_inferences":["A skeptical first check would be to run every model on exactly one shared test split (same filtered users, locations, and held-out weeks); the per-model dataset statistics in the paper's Table I suggest such a check could change the ranking.","Because the trajectory graph is built from training transitions only, the same hierarchical-convolution-plus-recent-attention design should apply to other sequential prediction problems where item transitions form a graph, such as web navigation or purchase sequences.","The 24-hour flow edge feature could be extended beyond week segmentation, for instance by conditioning the orientation module on time of day or adding hour prediction as a third task.","The authors say the implementation will be released after publication; if that release includes the exact preprocessed data split for each model, the comparability question can be settled directly."],"forward_implications":["If the reported gains hold on identical test splits, TrajGEOS is the current state of the art on the three standard LBSN benchmarks and a template for combining graph-learned location relations with multi-scale sequential preference.","The trajectory graph's edge features (distance, transition volume, hourly flow) can be injected directly into location embeddings, which suggests richer mobility counts belong in the representation rather than only in the loss.","The attention-based orientation module gives a mid-term preference signal at lower cost than a second recurrent network; replacing it with a GRU in the paper's own experiments is consistently worse.","The multi-task category head acts as a regularizer: the best location Recall@1 is found at alpha = 0.7, showing the auxiliary category loss can be balanced against the main task without degrading it.","The visual analyses indicate the model's advantage is spatially closer predictions and weaker for users with high location or category entropy, identifying where further gains are most needed."],"supporting_citations":[{"why":"Supplies the NYC and TKY check-in datasets used in the experiments.","marker":"[9]"},{"why":"Supplies the Dallas (Gowalla) check-in dataset used in the experiments.","marker":"[47]"},{"why":"A main baseline whose preprocessing and evaluation conventions the paper follows.","marker":"[17]"},{"why":"A long/short-term preference baseline that TrajGEOS directly extends and compares against.","marker":"[33]"},{"why":"Contributes the trajectory graph construction and location-embedding fusion reused in TrajGEOS.","marker":"[19]"},{"why":"The strongest reported baseline on NYC and Tokyo that the paper aims to surpass.","marker":"[43]"},{"why":"A geography-aware attention baseline included in the comparison tables.","marker":"[32]"},{"why":"Provides the edge-aware graph convolution used on the global trajectory graph.","marker":"[44]"},{"why":"Provides the neighborhood aggregation convolution used on user subgraphs.","marker":"[45]"},{"why":"Provides the periodic time encoding used to embed timestamps.","marker":"[46]"}],"fun_headline_variants":["TrajGEOS: Graph and orientation signals top next-location predictions","TrajGEOS: Hierarchical graph + orientation beats baselines in 3 cities","Graph-enhanced orientation network tops next-location Recall@1","TrajGEOS: Multi-level preferences lift next-location SOTA"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The ranking conclusion assumes the separately preprocessed datasets used for different baseline models are comparable enough to compare their scores, even though each model was run on its own subset of users, locations, and check-in records.","fun_headline_variants_meta":{"raw":{"variants":["TrajGEOS: Graph and orientation signals top next-location predictions","TrajGEOS: Hierarchical graph + orientation beats baselines in 3 cities","Graph-enhanced orientation network tops next-location Recall@1","TrajGEOS: Multi-level preferences lift next-location SOTA"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000718,"raw_usage":{"total_tokens":3267,"prompt_tokens":1026,"completion_tokens":2241,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":642,"completion_tokens_details":{"reasoning_tokens":2162}},"tokens_in":642,"tokens_out":2241,"duration_ms":17009,"temperature":1.0,"reasoning_tokens":2162,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:57:56.674171+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run every baseline and TrajGEOS on one identical user-level 80/20 split of the same filtered check-in data for NYC and Dallas, with identical test weeks and filtering rules, and check whether TrajGEOS still leads on Recall@1 and MRR@10 across five seeds; if the margin collapses, the reported outperformance is not established.","supporting_citations":[{"cited_title":"Modeling user activity preference by leveraging user spatial temporal characteristics in lbsns,","cited_arxiv_id":null,"evidence_quote":"Supplies the NYC and TKY check-in datasets used in the experiments."},{"cited_title":"Friendship and mobility: user movement in location-based social networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the Dallas (Gowalla) check-in dataset used in the experiments."},{"cited_title":"Deep- move: Predicting human mobility with attentional recurrent networks,","cited_arxiv_id":null,"evidence_quote":"A main baseline whose preprocessing and evaluation conventions the paper follows."},{"cited_title":"Where to go next: Modeling long-and short-term user preferences for point-of- interest recommendation,","cited_arxiv_id":null,"evidence_quote":"A long/short-term preference baseline that TrajGEOS directly extends and compares against."},{"cited_title":"Getnext: trajectory flow map enhanced transformer for next poi recommendation,","cited_arxiv_id":null,"evidence_quote":"Contributes the trajectory graph construction and location-embedding fusion reused in TrajGEOS."}],"review_version":1}