{"id":"8b7b632d-7c4a-4f34-b700-6360202b5cb5","arxiv_id":"2506.14070","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Using CaLLiPer, a contrastively trained location encoder that fuses coordinates and POI text, improves next-location prediction on some mobility datasets and shows the largest gains when test locations are unseen during training.","lead":"This paper applies an existing spatial-semantic location embedding model, CaLLiPer, to the task of predicting people's next visited place, on four public mobility datasets. It reports that this approach often beats older embedding methods, especially when the model must make predictions for places it has never seen during training.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Inductive protocol is underspecified: CaLLiPer may have pretrained on the held-out POIs while lookup-table baselines cannot embed unseen locations, biasing the central comparison.","rationale":"The paper is a useful empirical application of an existing model, with released code, four datasets, and an honest Section 5.1 that acknowledges CaLLiPer does not win everywhere. The central selling point, however, is the inductive setting, and that is exactly where the protocol is least secure. The manuscript explicitly claims L_new is unseen during pretraining, yet for the two Foursquare-based datasets the location set is defined by the same POI data CaLLiPer uses for contrastive pretraining, making leakage plausible unless an exclusion step is described and implemented. Separately, the baselines are lookup-table models with fixed vocabularies; the paper never explains how they produce embeddings for locations removed from their training sequences. Both problems are fixable, but until they are resolved the inductive comparison is not a fair test of CaLLiPer's claimed inductive advantage. The reader's weakest assumption identifies these same gaps, so I agree with that diagnosis and with the conditional recommendation; the requested clarification and ablation would settle whether the central claim survives.","tokens_in":15950,"tokens_out":3892,"duration_ms":44314,"concrete_test":"Run the inductive FSQ-NYC experiment twice: once with CaLLiPer's location encoder retrained after removing from its contrastive pretraining set all POIs whose locations fall in L_new, and once with lookup-table baselines given embeddings pretrained on the full mobility training set (so they have non-random vectors for L_new). Compare the inductive Acc@1 and MRR gaps against Table 2; if CaLLiPer's advantage shrinks substantially or reverses, the reported gap is an artifact of information leakage or unfairly weak baselines. Inspecting the released code to confirm the current handling of L_new in pretraining and downstream output layers is a necessary companion step.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The decisive issue is the inductive protocol in Section 4.3. The paper states that locations in L_new are 'not seen during neither the pre-training phase nor the downstream model training phase,' but for FSQ-NYC and FSQ-TKY the locations are the dataset's own POIs and CaLLiPer is pretrained on those same POI data (Sections 4.1 and A.1). Unless the POIs corresponding to L_new are explicitly removed from CaLLiPer's contrastive pretraining set, CaLLiPer has already seen the coordinates and textual descriptions of the test 'unseen' locations, so its inductive advantage is not evidence of generalisation. The same section is silent on how lookup-table baselines (Vanilla-E2E, Skip-gram, POI2Vec, Geo-Teaser, TALE, CTLE) obtain embeddings for L_new, which are absent from their fixed vocabularies after the sequence removal; if those embeddings are randomly initialised or the output layer is truncated, the comparison is biased from the outset. Without stating and enforcing an equal-information condition for all methods, the inductive column of Table 2 does not establish the central claim of superior generalisation to previously unseen locations.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes using CaLLiPer, a contrastive location encoder pre-trained on POI coordinates and textual descriptions, as an inductive location embedding for individual next-location prediction. On four public mobility datasets, it compares CaLLiPer with six baselines under conventional and inductive train/validation/test splits, using MHSA as a common downstream predictor, and reports Acc@1/5/10, MRR, and nDCG@10 in Table 2, together with UMAP visualisations. The abstract claims consistent superiority, with a particular advantage in inductive scenarios involving previously unseen locations, and the authors release code and data for reproducibility.","tokens_in":16335,"tokens_out":5669,"duration_ms":54161,"significance":"If the central claim were established, this would be a useful contribution: it would be the first application of a multimodal, spatially explicit and semantically informed location encoder to individual mobility prediction, and it would offer a practical way to handle emerging locations without retraining lookup-table embeddings. The manuscript is transparent about releasing code and data, and the evaluation uses public datasets with a common downstream model. However, the inductive protocol is currently underspecified in a way that directly affects the validity of the main claim, and the abstract overstates the results reported in Table 2. The significance is therefore conditional on a corrected, equal-information inductive evaluation.","major_comments":[{"comment":"The inductive setting does not enforce an equal-information condition. Section 4.3 removes mobility sequences containing L_new from the train/validation sets and states that locations in L_new are 'not seen during neither the pre-training phase nor the downstream model training phase.' However, CaLLiPer's pre-training is not on mobility sequences but on POI coordinates and textual descriptions (Sections 3.3.1 and A.1). For FSQ-NYC and FSQ-TKY the locations are the dataset's own POIs, and Section 4.1 states that no additional POI sourcing is needed because these datasets already include POI information; for Gowalla-LD and Geolife, external Foursquare POI data covering London and Beijing are used for CaLLiPer pre-training. The paper never states that the POIs corresponding to L_new are removed from CaLLiPer's contrastive pre-training set. If they are not removed, CaLLiPer has already seen the coordinates and textual descriptions of the supposedly unseen test locations, and the inductive comparison measures data leakage rather than generalisation.","section":"4.3 (with 4.1 and A.1)"},{"comment":"The paper does not explain how the non-inductive baselines obtain embeddings for locations in L_new. Vanilla-E2E, Skip-gram, POI2Vec, Geo-Teaser, TALE, and CTLE all rely on a fixed vocabulary learned from the training sequences; after the inductive filtering, no sequence containing an L_new location remains, so these methods have no embedding for those locations. The downstream model's output layer, defined in Eq. (4) and trained with Eq. (5), has one logit per location, and the paper does not specify whether L_new are in the output space for the baselines, whether their embeddings are randomly initialised, or whether the final layer is truncated. Without this information, the inductive columns of Table 2 are difficult to interpret as a comparison of generalisation ability.","section":"4.3 (with Eqs. 4-5)"},{"comment":"The abstract's claim that CaLLiPer 'consistently outperforms strong baselines' and 'particularly excelling in inductive scenarios' is not supported by Table 2. On FSQ-TKY conventional, the relative differences are negative for all five metrics (Acc@1 -8.31%, Acc@5 -1.09%, Acc@10 -0.30%, MRR -3.52%, nDCG@10 -2.47%); on Geolife, Acc@5 and Acc@10 are negative in both conventional and inductive settings. The more careful wording in Section 5.1, which says CaLLiPer has the best result in the majority of cases and performs well on a subset of metrics in the remaining datasets, is accurate, but it should replace the abstract's overclaim.","section":"Abstract and Section 5.1 / Table 2"}],"minor_comments":[{"comment":"The sentence 'not seen during neither the pre-training phase nor the downstream model training phase' should read 'not seen during either the pre-training phase or the downstream model training phase.'","section":"4.3"},{"comment":"Both figure captions contain the typo 'conventinoal setting,' which should be corrected to 'conventional setting.'","section":"Figures 4 and 5"},{"comment":"The dataset enumeration lists only FSQ-NYC, Gowalla-LD, and Geolife, but Table 1 and the rest of the paper also use FSQ-TKY; the fourth dataset should be named explicitly in the text.","section":"4.1"},{"comment":"The phrase 'significantly outperforms' is used without reporting significance tests; several key comparisons show overlapping standard deviations, for example FSQ-NYC inductive Acc@1 with CaLLiPer 16.07 ± 0.69 versus Vanilla-E2E 15.25 ± 1.04, so either significance testing or explicit wording that the differences are not tested would be appropriate.","section":"5.1"},{"comment":"Appendix A.1 reports dataset-specific hyperparameters for CaLLiPer, but the accompanying text should be cross-checked for FSQ-TKY, which is omitted from the dataset list in Section 4.1.","section":"A.1"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the decisive issue is the inductive protocol. If the authors can demonstrate, by rerunning CaLLiPer pre-training with the held-out POIs removed or by explicitly stating that this was already done, and if they specify how baselines handle unseen locations, then a revised version could be acceptable. The abstract's overclaim relative to Table 2 should also be corrected. I do not see a circularity problem beyond the data-leakage concern raised above."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, you should know two things: the paper is a legitimate application study of a strong prior model, and its headline inductive result is undercut by a protocol gap that is easy to miss.\n\nWhat's new: the authors take CaLLiPer, their earlier POI-contrastive location encoder, and apply it as the embedding layer for next-location prediction. That application is new, and so is the inductive split they design to simulate unseen locations. They run four public datasets, use a common downstream model, release code, and the conventional-setting results are reasonably strong: CaLLiPer is best on all metrics on FSQ-NYC and Gowalla-LD, and is competitive on the other two. That is real evidence that POI-based spatial-semantic embeddings help for mobility, and the code will be useful to the community.\n\nThe soft spot is the inductive protocol. Section 4.3 defines L_new as locations whose mobility sequences are removed from training and validation, but it never says what happens to the embeddings of those locations. For CaLLiPer, pretraining uses the POI data from these same datasets; for FSQ-NYC and FSQ-TKY the POIs are the dataset's own locations, so the held-out L_new POIs were almost certainly fed into CaLLiPer's contrastive pretraining. Calling them 'unseen' is therefore misleading. For the baselines, all six are lookup-table models with fixed vocabularies; after removing sequences they have no way to produce vectors for L_new. The paper is silent on whether they get random, zero, or re-normalized vectors. That asymmetry alone could explain much of CaLLiPer's inductive advantage. This is not a minor detail; it is the load-bearing claim of the paper.\n\nThe abstract also overstates the results. Table 2 shows CaLLiPer loses on FSQ-TKY conventional and is mixed on Geolife, so 'consistently outperforms' is not accurate. The body is more careful, but the abstract and conclusion are not.\n\nBottom line: as a conventional embedding comparison, the paper is useful and mostly sound. As a demonstration of generalization to unseen locations, it does not yet establish the claim. I would not cite the inductive numbers until the protocol is fixed or the claim is reframed. For peer review, I'd send it out — the idea is timely and the benchmark is valuable — but I'd expect substantial revision: exclude the held-out POIs from CaLLiPer's pretraining, specify how baselines handle out-of-vocabulary locations, and tone down the abstract.","headline":"A useful application study whose inductive claim is not yet established: the held-out POIs are likely not held out from CaLLiPer's pretraining, and baselines are left without a mechanism to embed unseen locations.","tokens_in":16729,"tokens_out":3683,"would_cite":false,"duration_ms":32869,"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":"The paper claims that CaLLiPer, a multimodal location embedding learned by aligning coordinates with point-of-interest text through contrastive learning, improves next-location prediction, and that its main advantage appears when the…","keywords":["human mobility","next location prediction","location embedding","points of interest","contrastive learning","inductive learning","spatial semantics","location-based services"],"falsifier":"Re-run the FSQ-NYC inductive experiment after deleting from CaLLiPer's pre-training corpus every POI whose coordinates overlap the held-out set L_new, while keeping the downstream training identical. If the Acc@1 advantage over the best baseline (reportedly 16.07 versus 15.25) disappears, the claimed generalisation to unseen places fails the 'unseen' test.","tokens_in":15770,"feed_emoji":"📍","tokens_out":6500,"duration_ms":62251,"temperature":0.7,"pith_summary":"The paper tries to show that location embeddings for next-location prediction need not be learned from historical mobility traces alone. It applies CaLLiPer, a contrastively trained multimodal encoder that maps coordinates and point-of-interest (POI) text into a shared space, as a drop-in replacement for the lookup-table embeddings used by most predictors. Experiments on four mobility datasets compare this embedding with six baselines under a conventional split and under an inductive split that hides ten percent of locations from training. The claim is that CaLLiPer matches or beats the baselines in most settings, and that its advantage is largest when the test set contains locations the predictor has never seen. If true, it means predictors can keep working when users visit new places without retraining the embedding layer.","feed_headline":"POI-trained location embeddings beat baselines on unseen places","feed_subtitle":"CaLLiPer fuses coordinates and place semantics so next-location predictors keep working when new places appear.","key_machinery":"The load-bearing object is CaLLiPer, a multimodal representation-learning framework that aligns a coordinate-based location encoder with a text encoder over point-of-interest descriptions. The location encoder maps geographic coordinates through a multi-scale Grid positional encoding and a fully connected network; the text encoder is a Sentence Transformer; and the two modalities are pulled together by a bidirectional InfoNCE contrastive objective. Because the encoder maps any coordinate rather than only locations in a fixed vocabulary, it is inductive: an embedding exists for any unseen location. The downstream mechanism is MHSA, a multi-head self-attention Transformer that ingests these frozen embeddings and outputs a softmax over candidate next locations; the paper uses its performance as the measurement of embedding quality.","core_discovery":"The central discovery is that location embeddings built from general spatial and semantic data, rather than from individual mobility traces, can serve as drop-in substitutes for conventional lookup-table embeddings in next-location prediction. The paper applies CaLLiPer, pre-trained by contrasting coordinates with accompanying point-of-interest text, to four public mobility datasets in both a conventional setting and an inductive setting where ten percent of locations are absent from the training and validation sequences. The results show CaLLiPer attains the best scores on all metrics in both settings for FSQ-NYC and Gowalla-LD; on FSQ-TKY it becomes best on most metrics in the inductive setting; and on Geolife it is best on three of five metrics in both settings. From these numbers the paper concludes that the inductive, spatially explicit, semantically enriched nature of CaLLiPer embeddings makes them particularly suited to scenarios where new locations emerge.","pith_inferences":["A stronger inductive test would also withhold the held-out locations' POI descriptions from CaLLiPer's pre-training corpus; the paper does not report whether such overlap exists, so part of the 'unseen' advantage could come from the semantic channel rather than from spatial continuity alone.","Comparing CaLLiPer against a coordinate-only location encoder (without POI text) would separate the contribution of spatial continuity from semantic enrichment, an ablation the paper does not run.","The same embeddings could be tested for cold-start mobility prediction, where a new user's first check-in is to an unseen location; if the inductive mechanism works, the predictor should degrade far less for users with few observations than lookup-table baselines."],"forward_implications":["Predictive systems can serve new locations without retraining the embedding layer, because the location encoder produces vectors for any coordinate at inference time.","Embedding pre-training no longer depends on mobility traces, so it can exploit publicly available POI data even when detailed movement data are privacy-restricted.","On two datasets the inductive gains are larger than the conventional gains, and on a third the method moves from behind in the conventional setting to leading most metrics in the inductive setting, indicating the spatial-semantic encoder carries much of the generalisation load.","The same frozen embeddings feed the same downstream predictor across datasets and settings, so the representation can act as reusable infrastructure rather than a task-specific lookup table."],"supporting_citations":[{"why":"Supplies the CaLLiPer model that this paper applies to location embedding.","marker":"[34]"},{"why":"Provides the MHSA downstream predictor and the seven-day sliding-window preprocessing convention.","marker":"[11]"},{"why":"Provides the Word2Vec basis for the Skip-gram location embedding baseline.","marker":"[26]"},{"why":"Contributes POI2Vec, a spatial Word2Vec baseline that must be beaten.","marker":"[8]"},{"why":"Contributes CTLE, a BERT-based context- and time-aware location embedding baseline.","marker":"[16]"},{"why":"Provides the FSQ-NYC check-in dataset used in the main comparison.","marker":"[36]"},{"why":"Provides the Gowalla-LD check-in dataset.","marker":"[4]"},{"why":"Provides the Geolife GNSS trajectory dataset.","marker":"[40]"},{"why":"Supplies external Foursquare POI text needed to train CaLLiPer for Gowalla-LD and Geolife.","marker":"[9]"},{"why":"Provides the Sentence Transformer text encoder inside CaLLiPer.","marker":"[28]"}],"fun_headline_variants":["CaLLiPer beats baselines for unseen locations","Spatial-semantic embeddings predict mobility beyond visited places","Inductive location embeddings outperform on new places","Fusing coordinates and POI semantics boosts next-location prediction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The inductive comparison is only fair if every method gets the same kind of information about unseen locations; the paper leaves open whether baselines receive only random or zero vectors for held-out locations while CaLLiPer still benefits from their coordinates and POI semantics.","fun_headline_variants_meta":{"raw":{"variants":["CaLLiPer beats baselines for unseen locations","Spatial-semantic embeddings predict mobility beyond visited places","Inductive location embeddings outperform on new places","Fusing coordinates and POI semantics boosts next-location prediction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00086,"raw_usage":{"total_tokens":3724,"prompt_tokens":927,"completion_tokens":2797,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":543,"completion_tokens_details":{"reasoning_tokens":2733}},"tokens_in":543,"tokens_out":2797,"duration_ms":22761,"temperature":1.0,"reasoning_tokens":2733,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:53:35.981175+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the FSQ-NYC inductive experiment after deleting from CaLLiPer's pre-training corpus every POI whose coordinates overlap the held-out set L_new, while keeping the downstream training identical. If the Acc@1 advantage over the best baseline (reportedly 16.07 versus 15.25) disappears, the claimed generalisation to unseen places fails the 'unseen' test.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the MHSA downstream predictor and the seven-day sliding-window preprocessing convention."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Geolife GNSS trajectory dataset."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies external Foursquare POI text needed to train CaLLiPer for Gowalla-LD and Geolife."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Sentence Transformer text encoder inside CaLLiPer."}],"review_version":1}