{"id":"aac306c1-504e-4aae-bcac-d701db281765","arxiv_id":"2506.21612","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"AdaptGOT combines four graph sampling strategies, a geography-co-occurrence-text attention module, and a mixture-of-experts aggregator to produce task-adaptive POI embeddings that outperform baseline pre-trained embeddings in the reported experiments.","lead":"This paper introduces AdaptGOT, a pre-trained model that represents points of interest using geographic location, co-occurrence, and text, and adapts those representations to different downstream tasks. The authors report higher recall than several pre-trained baselines on next-POI, category, and POI recommendation tasks using Foursquare and Yelp data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 4.4's adaptive aggregator cannot be task-adaptive as claimed: Eqs. 20-22 contain no task input, and Eq. 23's JS divergence is never defined, so the central 'adaptive across tasks without retraining' contribution is unsupported.","rationale":"The paper's headline novelty is adaptive representation learning, not merely another fixed embedding. The reader flagged Eq. 23 as under-specified; I agree, and a close reading of Eqs. 20-22 goes further: even a well-defined JS loss would not make the embedding task-adaptive at inference, because no task variable exists in the forward path. The central empirical claim may survive as a claim about a strong fixed embedding, but the 'adaptive' part of the title and the Gap 3 contribution would be unsupported. This is a correctness and novelty issue rather than a disagreement with field consensus, and it is addressable: the authors could condition the gate on a task embedding, clarify how task adaptivity occurs, or retract the without-retraining claim. Since the empirical results are not otherwise falsified, the reader's CONDITIONAL verdict is appropriate; no change is needed, but the revision conditions should include exposing the code, demonstrating task-conditioning, and specifying how L_JS is computed and connected to the forward pass.","tokens_in":19614,"tokens_out":12013,"duration_ms":128013,"concrete_test":"Obtain the linked anonymous training code and check two things: (i) whether any task-identifying input or downstream gradient reaches Eqs. 20-22 at inference; (ii) whether L_JS is computed with an explicit estimator of P(G_s) and P(G) and is back-propagated into h'_i. Then run one ablation: remove L_JS and re-measure the Table 2 rows for FPMC and GETNEXT on NY and LA at Rec@5/10/15. If h'_i is identical across tasks for the same POI, or if removing L_JS leaves all three metrics unchanged within seed noise, the 'adaptive across tasks' and 'topological consistency' claims are not supported by the described model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 1 and Section 4.4 claim that the MoE aggregator dynamically adjusts POI representations across tasks without retraining. However, the forward pass in Eqs. 20-22 computes h'_i only from the POI's GOT-attention output and learned gating weights; no task identifier, task loss, or downstream query enters these equations. The same pre-trained POI therefore receives exactly the same embedding for next-POI recommendation, category prediction, and POI recommendation unless the downstream model updates the embedding, which is precisely the retraining the paper says is unnecessary. Separately, the JS-divergence term claimed to ensure topological consistency (Eq. 23) is not operationalized: P(G_s) and P(G) are never defined or estimated, and no equation links this loss to h'_i or to any graph-topological quantity. As written, the adaptive/topological contribution reduces to a fixed per-node MoE-weighted embedding, so the distinctiveness of AdaptGOT over a fixed pre-trained embedding is unverified. This does not disprove the empirical superiority claims in Table 2, but it removes the paper's stated mechanism for addressing Gap 3.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AdaptGOT, a pre-training framework for POI embeddings. It combines mixed subgraph sampling (KNN, density, importance, and category-aware strategies), a GOT representation module that fuses geographical, co-occurrence, and textual information, a modified GOT attention mechanism, and an MoE-based 'adaptive representation aggregator' claimed to adjust POI representations across downstream tasks without retraining by minimizing a Jensen-Shannon divergence between subgraph and original graph distributions. The model is trained self-supervised with text, geography, and co-occurrence reconstruction losses. Experiments on Foursquare and Yelp evaluate next-POI recommendation, next-category prediction, and POI recommendation, with ablations, sensitivity analyses, and a zero-shot cross-city transfer test; the paper reports consistent recall improvements over SpaBERT, CTLE, and several other baselines.","tokens_in":19895,"tokens_out":6010,"duration_ms":60439,"significance":"If its mechanism claims were fully supported, AdaptGOT would be a useful contribution to POI representation learning: the multi-context sampling and GOT attention are sensible design choices, and the evaluation breadth (multiple datasets, tasks, ablations, cross-city transfer) is a strength. However, the central 'adaptive across tasks without retraining' claim is not implemented in the forward pass as written, and the JS-divergence objective is not operationalized. The empirical comparison also lacks variance reporting and selects hyperparameters on the evaluation metrics. The contribution is therefore currently under-supported, though these issues do not by themselves disprove the reported recall gains.","major_comments":[{"comment":"The forward pass in Eqs. (20)–(22) computes h'_i from the GOT-attention output z_i and learned gating weights only; no task identifier, task loss, or downstream query appears in these equations. This contradicts the claim in Section 1 and Section 4.4 that the MoE aggregator 'dynamically adjusts POI representations across tasks without retraining': a fixed pretrained POI receives the same h'_i for next-POI recommendation, category prediction, and POI recommendation unless the downstream model fine-tunes the embedding. Please state precisely what 'adaptive across tasks' means operationally, or revise the claim to describe a fixed multimodal embedding.","section":"§4.4, Eqs. (20)–(22)"},{"comment":"The JS-divergence term that is supposed to preserve 'topological consistency' is not operationalized. P(G_s) and P(G) are never defined or estimated, no equation links D_JS to h'_i or to a graph-topological quantity, and the claim that minimizing Eq. (23) aligns subgraph representations with global structure is therefore unsupported. As written, the adaptive representation aggregator reduces to a learned per-node gating network, and the JS loss contributes nothing identifiable to the mechanism. Define the distributions, give an estimator, and show how the loss affects the gating or expert outputs.","section":"§4.4, Eq. (23)"},{"comment":"The theoretical claims supporting the 'expressive power' contribution are not established in the manuscript. The proofs are deferred to an anonymous external README, and the statements in Eqs. (4)–(6) are not formal: P_conflict and entropy are not defined on the same probability space, and the proof sketch ('feature space size grows exponentially') does not by itself imply the claimed 1-WL discriminative-power result. Either provide complete proofs in the paper or reframe these statements as intuition and empirical motivation.","section":"§3.2, Theorems 1–3"},{"comment":"The empirical evidence for superiority is weakened by the absence of error bars and significance tests, and by the fact that key hyperparameters (number of subgraphs, heads, neighbors) are selected on the same Recall@K metrics used for the final comparison. Table 2 also contains cases where the advertised gain is zero or negligible, for example LA-Next POI Rec@5 under LSTPM: GOT 0.078 versus SpaBERT 0.078. Please report means over repeated runs with variance, use a validation-based selection protocol, and identify where the improvement is statistically reliable.","section":"§5.1.4 and Table 2"}],"minor_comments":[{"comment":"The term 'GOT' is used ambiguously for both the model (AdaptGOT) and the representation module (GOT representation); Table 2's 'GOT' row should be labeled consistently to avoid confusion.","section":"Throughout"},{"comment":"There are numerous typos and formatting artifacts: 'POl' in the Abstract, 'onGeographical location' in the Abstract, the repeated 'the' in Section 4.4, and the placeholder ACM template with 2018 copyright and 'Conference acronym ’XX'.","section":"Abstract and template"},{"comment":"The text refers to 'Los Angeles' when the dataset column is labeled 'LA' for Louisiana; please make the city/state labels consistent.","section":"§5.2.1"},{"comment":"Equation (28) uses the notation g ∈ G(h_i) without defining the distribution over gates, and Eq. (8) uses N_den both as a set and as its cardinality; please clean up the notation.","section":"Eqs. (8) and (28)"},{"comment":"The footnote pointing to an anonymous README for the proofs and NDCG results is not an adequate substitute for material that should be in the paper or a stable repository; include the proofs and full metric tables in the manuscript.","section":"Footnote 1"},{"comment":"The zero-shot cross-city table does not state which source/target city is used for each column; without this information, the transfer claim is hard to interpret.","section":"Table 5"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the stress-test concern is substantive. The 'adaptive without retraining' mechanism is not present in the forward pass of Eqs. (20)–(22), and the JS divergence in Eq. (23) is not operationalized. If the authors cannot address this, the paper's distinctive contribution over fixed pretrained embeddings is reduced to the empirical gains, which themselves need stronger statistical support. The manuscript also appears to be a raw preprint with placeholder formatting; it should be brought up to the venue's submission standards before review."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the empirical package is worth engaging with, but the paper's headline mechanism—task-adaptive embeddings without retraining—does not at present follow from its own equations. The stress-test note is right: Eqs. 20–22 take no task input, so the MoE gating produces the same h'_i for every downstream task unless the downstream model fine-tunes, which is exactly the retraining the paper says is unnecessary. That should be fixed or the claim toned down.\n\nWhat's new: AdaptGOT is genuinely a different combination—four subgraph sampling strategies (KNN, density, importance, category) feeding a geo/co-occurrence/text attention encoder, with an MoE aggregator over the subgraph embeddings. I don't know prior work with that specific mix, and the two-dataset, multiple-downstream-model evaluation is respectable. The ablation tables (Fig. 3, Table 3) show each component matters, with density and text especially impactful. Those results are plausible.\n\nThe soft spots, in rough order of severity:\n\n1. The JS-divergence term in Eq. 23 is not operationalized. P(G_s) and P(G) are never defined, no estimator is given, and the midpoint M is written with G_g, which is inconsistent with the rest of the text. As written, the “topological consistency” loss could be anything. This is not a cosmetic issue if it is claimed as a mechanism.\n\n2. The “adaptive across tasks” claim is unsupported for the reason above. What the MoE actually does is learn a per-node weighting over fixed subgraph embeddings. That is still useful, but it is not task-adaptive.\n\n3. No error bars or significance tests. The grid search in Sec. 5.1.4 tunes on the same metrics used for reporting, so some of the gains may be optimistic.\n\n4. Theorems 1–3 are informal restatements of standard 1-WL/entropy intuition, supported by a toy example rather than a proof. Fine as motivation; not a paper contribution.\n\n5. Missing comparisons: MGeo and M3PT are discussed but not in Table 2.\n\nThe citation pattern is fine. The writing is rough in places, but the experiments are detailed enough that a careful referee could reproduce them.\n\nWho this is for: anyone working on POI embeddings, geospatial pre-training, or context-aware GNN recommender systems. It deserves a serious referee—the empirical architecture is novel and the negative result on the adaptivity claim is a good finding. I'd send it to a strong reviewer, but insist the authors either condition the MoE on task information or drop the “without retraining” language. That's likely a major revision, not a reject.","headline":"Empirical POI embedding work that deserves a review, but the headline 'task-adaptive without retraining' is not backed by the equations, and the JS-divergence loss is underspecified.","tokens_in":20404,"tokens_out":2817,"would_cite":true,"duration_ms":26858,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Adaptive POI embeddings beat fixed pretrained baselines by up to 23.8%","keywords":["point-of-interest embedding","pretrained representation learning","mixture of experts","graph attention","subgraph sampling","next POI recommendation","Jensen-Shannon divergence","geo-contextual learning"],"falsifier":"Run AdaptGOT with the JS divergence loss replaced by a fixed constant, keeping everything else identical, and compare the learned MoE gate weights and downstream Recall@K. If the gate weights and recall are unchanged, the JS term is not carrying the topological-consistency mechanism; conversely, if the term measurably changes which experts are selected, the claimed mechanism has a verifiable effect. A second check is to define $P(G_s)$ and $P(G)$ explicitly on a small POI graph and recompute Eq. (23) to see whether the minimized value actually tracks topological differences.","tokens_in":19414,"feed_emoji":"📍","tokens_out":5777,"duration_ms":54687,"temperature":0.7,"pith_summary":"AdaptGOT is a pre-trained model for point-of-interest (POI) embeddings that claims a single embedding can serve multiple downstream tasks by adapting itself per task. The model fuses three contexts — geographical location, co-occurrence patterns, and user-generated text — and combines four subgraph sampling strategies so that neighborhoods reflect proximity, density, importance, and category. A mixture-of-experts aggregator then weights these contextual subgraphs differently for each task without retraining, while a Jensen-Shannon divergence term is meant to keep subgraph representations topologically faithful to the full POI graph. On Foursquare (New York, Tokyo) and Yelp (Louisiana, Nevada) data, the paper reports consistent Recall@K improvements over fixed pretrained embeddings across next-POI recommendation, next-category prediction, and POI recommendation, with the largest stated gain being a 23.8% Rec@5 improvement over SpaBERT under FPMC in New York. A sympathetic reader would take the central thesis to be that task-adaptive, multi-context pretraining beats fixed one-size-fits-all POI embeddings.","feed_headline":"Adaptive POI embeddings beat fixed pretrained baselines by up to 23.8%","feed_subtitle":"A pre-trained model mixes geography, co-occurrence, and text, then switches context experts per task for better next-POI recall.","key_machinery":"The load-bearing mechanism is the adaptive encoder–decoder built on a sparse mixture-of-experts (MoE) layer. Given four candidate subgraphs produced by KNN, density-based, importance, and category-aware sampling, the MoE gating function computes a sparse, task-dependent weighting, and the model minimizes the Jensen–Shannon divergence between the weighted subgraph feature distribution and the original graph feature distribution to preserve topological consistency. Around this sit the GOT attention module, which folds relative geographical position and normalized co-occurrence into the attention score while BERT-encoded review text forms node features, and a masked self-supervised objective that reconstructs text, geo, and co-occurrence signals. The expressive-power argument (Theorems 1–3) ties the design to the 1-WL isomorphism test: multi-context features reduce label conflicts and increase entropy, which is why mixed subgraphs are claimed to beat single-context message passing.","core_discovery":"On its own terms, AdaptGOT claims to establish that pre-trained POI representations are not required to be fixed. By jointly encoding geographical, co-occurrence, and textual signals in a modified graph attention mechanism and letting a sparse mixture-of-experts layer pick different subgraph topologies for different tasks, the model produces embeddings that outperform fixed pretrained baselines (CTLE, SpaBERT) and random or fixed initializations across three POI tasks on two real-world datasets. The paper also claims theoretical support: multiple contexts and multiple subgraphs increase the discriminative power and entropy of node embeddings under the 1-WL isomorphism test, which is offered as the reason mixed sampling improves expressiveness. Ablations attribute the gains to each component — removing the adaptive aggregator, the attention module, or any single context degrades recall — and a zero-shot cross-city experiment reports an average recall increase of 19.7%.","pith_inferences":["If the JS-divergence term turns out not to be well-defined, the practical contribution likely survives as a task-adaptive gating network over multi-context subgraphs; the loss could be replaced by a simpler diversity or entropy regularizer.","The same mixed-sampling-plus-MoE recipe could transfer to other geo-entity problems such as land-use classification, region-function identification, or route prediction, where no single neighborhood definition is correct for all queries.","A direct test of the 1-WL argument would be to measure embedding entropy or oversmoothing (for example, average cosine similarity between neighboring node embeddings) across subgraph counts; the theory predicts less oversmoothing with more contexts."],"forward_implications":["One pre-trained POI embedding can be adapted at inference time to tasks with different contextual emphases, so downstream models no longer need task-specific fine-tuning of the embedding layer.","Cross-city transfer improves: in the paper's zero-shot experiment, AdaptGOT raises average recall by 19.7% over the fixed baselines, suggesting embeddings trained on one city can serve another.","Each context channel earns its keep: ablations show removing geography, co-occurrence, text, or the adaptive aggregator lowers recall, so multi-context integration is doing measurable work.","The expressive-power argument implies that adding more POI contexts or more subgraphs should keep improving label discrimination up to a point, with diminishing returns after four subgraphs as observed."],"supporting_citations":[{"why":"The fixed geo-entity embedding baseline (SpaBERT) that AdaptGOT must beat on all three tasks; removing it removes the central comparison.","marker":"[22]"},{"why":"The pre-trained trajectory-context location embedding baseline (CTLE) that AdaptGOT improves on, especially in next-POI recommendation.","marker":"[24]"},{"why":"The FPMC downstream model where the paper reports its largest gains, including the 23.8% Rec@5 improvement over SpaBERT.","marker":"[31]"},{"why":"The mixture-of-experts literature that motivates the adaptive aggregator's sparse gating over contextual subgraphs.","marker":"[28]"},{"why":"The graph mixture-of-experts method supplying the noisy-topk gating formulation and the importance loss for balancing experts.","marker":"[35]"},{"why":"The 1-WL isomorphism test used as the theoretical lens for why multiple contexts and subgraphs increase expressive power.","marker":"[29]"},{"why":"The 'from stars to subgraphs' result that grounds the claim that subgraph-based embeddings can surpass message-passing expressiveness.","marker":"[49]"},{"why":"The BERT encoder used to embed mixed user review text as the textual node feature in the GOT representation.","marker":"[10]"},{"why":"The neighbor-sampling baseline (BNS) used in the case study to contrast subgraph quality and over-smoothing.","marker":"[43]"}],"fun_headline_variants":["AdaptGOT: adaptive POI embeddings that switch experts per task","Pre-trained POI model fuses geo, text, and co-occurrence adaptively","Mixture-of-experts POI embeddings beat fixed pretrained baselines","Zero-shot POI recall improves 19.7% with context-aware pretraining","AdaptGOT blends geography, co-occurrence, and text for POI tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole 'adaptive and topologically consistent' claim rests on the assumption that the Jensen–Shannon divergence in Eq. (23) is a well-defined objective between identifiable distributions of subgraph and original-graph features, yet the paper never defines how $P(G_s)$ and $P(G)$ are computed; if that term is not well-defined, the adaptive module reduces to a learned gating network and the stated mechanism is not doing the claimed work.","fun_headline_variants_meta":{"raw":{"variants":["AdaptGOT: adaptive POI embeddings that switch experts per task","Pre-trained POI model fuses geo, text, and co-occurrence adaptively","Mixture-of-experts POI embeddings beat fixed pretrained baselines","Zero-shot POI recall improves 19.7% with context-aware pretraining","AdaptGOT blends geography, co-occurrence, and text for POI tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000288,"raw_usage":{"total_tokens":1707,"prompt_tokens":982,"completion_tokens":725,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":598,"completion_tokens_details":{"reasoning_tokens":623}},"tokens_in":598,"tokens_out":725,"duration_ms":6925,"temperature":1.0,"reasoning_tokens":623,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:05:47.561592+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run AdaptGOT with the JS divergence loss replaced by a fixed constant, keeping everything else identical, and compare the learned MoE gate weights and downstream Recall@K. If the gate weights and recall are unchanged, the JS term is not carrying the topological-consistency mechanism; conversely, if the term measurably changes which experts are selected, the claimed mechanism has a verifiable effect. A second check is to define $P(G_s)$ and $P(G)$ explicitly on a small POI graph and recompute Eq. (23) to see whether the minimized value actually tracks topological differences.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The pre-trained trajectory-context location embedding baseline (CTLE) that AdaptGOT improves on, especially in next-POI recommendation."},{"cited_title":"Graph Mixture of Experts: Learning on Large-Scale Graphs with Explicit Diversity Modeling","cited_arxiv_id":"2304.02806","evidence_quote":"The graph mixture-of-experts method supplying the noisy-topk gating formulation and the importance loss for balancing experts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The 1-WL isomorphism test used as the theoretical lens for why multiple contexts and subgraphs increase expressive power."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The neighbor-sampling baseline (BNS) used in the case study to contrast subgraph quality and over-smoothing."}],"review_version":2}