acceptodds
Under review as a conference paper at ICLR 2027

Taming Continuous Location Tokens for Generalized Next Point-of-Interest Recommendation

Abstract

Next point-of-interest (POI) prediction is the task of estimating a user’s future geographical location, supporting applications from restaurant recommendation to real estate shopping. Many state-of-the-art methods assign discrete embeddings to each POI, ignoring spatial relationships and limiting knowledge transfer to nearby unseen locations. To address this limitation, our framework combines a continuous latitude–longitude encoder using random Fourier features (RFF) with a causal transformer that predicts continuous location representations from historical check-in sequences. On Massive-STEPS, a multi-city check-in dataset, we examine how RFF frequency bandwidth controls representation granularity: lower frequencies enforce smooth spatial priors, whereas higher frequencies encourage precise coordinate memorization. We illustrate this tradeoff by withholding a portion of locations from training and comparing the recall on both seen and unseen POIs: a lower-frequency RFF configuration raises average unseen Recall@100 across six city datasets from 36.4% to 42.9%, while seen recall falls from 63.8% to 57.5%. Further lowering the frequencies does not consistently improve unseen recall, indicating an optimal granularity among the tested settings. Additionally, we compare RFF with discrete location encoders and use a national-scale real estate shopping dataset to validate our approach. These findings provide empirical guidance for choosing the spatial scale of RFF embeddings so that learned geographic patterns generalize to destinations absent from training data.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
POI prediction, Random Fourier Features, Location prediction, Location representation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_rsmpOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies generalized next-POI recommendation, focusing on locations that are unseen during training. The authors use Random Fourier Features to encode continuous latitude–longitude coordinates into location representations and combine them with a causal Transformer for next-location prediction. The approach is evaluated on Massive-STEPS and a national-scale real-estate dataset to show its effectiveness.

Critical strengths

S1: Recommendation for newly introduced POIs and cold-start locations is a practical problem.

S2: The paper is easyto read and follow.

Critical weaknesses

W1: One degree of longitude corresponds to substantially different physical distances at different latitudes. RFF operates on the standardized coordinate space rather than directly on geodesic distance. Thus I am afraid the paper's treatment of latitude and longitude is not reasonable. Comparison with other treatment is needed.

W2: The novelty is limited. Most of the main components is the embedding and tricks for this embedding.

W3: The benefit of the learned location encoder over heuristic prior is weak. . In Table 3, Gravity+history already achieves an average unseen R@100 of 42.2, whereas the best reported RFF setting reaches merely 42.9.

W4: Analysis in Table 3 is evaluated primarily with R@100. This is a relatively permissive metric given that the candidate sets can contain tens of thousands of locations. The paper should additionally report smaller metrics such as R@5, R@10, R@20, and ranking-sensitive metrics such as MRR or NDCG.

W5: Code is not released, thus hindering reproductivity.

W6: More analysis of computational efficiency and deployment cost is needed.

Questions for authors

Q1: A new location may contain POIs from different categories. Is a single location embedding sufficient to represent them? Moreover, POI semantics are not completely determined by location. For example, a newly built shopping mall may contain many POIs belonging to very different categories, and an existing POI may change its category over time. Would these cases be considered new POIs under the paper's definition? If so, the current coordinate-based embedding seems unable to capture such semantic differences.

Q2: The paper creates training-unseen locations by dropping 20% POIs. It is unclear whether performance on the resulting split reflects generalization to naturally emerging new POIs. (or, whether such dropping reflects the real cases).

4 comments

We thank the reviewer for the detailed technical review. We address each of the weaknesses below:

W1 (lat/lon degrees vs. geodesic distance). We agree that one degree of longitude spans different distances at different latitudes, and we note two mitigations, one already in the paper and one new experiment:

(i) As detailed in Appendix B, we standardize latitude and longitude independently for each city using training statistics. Within a single metropolitan area, the longitudinal compression factor (cos⁡(latitude)\cos(\text{latitude})) is nearly constant, varying by less than 0.4% across New York.

(ii) As suggested by the reviewer, we conduct new experiments with two geodesic-aware treatments, both on New York and Istanbul under the same protocol and with a selection of different frequency parameters: (a) RFF with 3D Cartesian coordinates on the unit sphere (x, y, z from lat/lon); and (b) Sphere2Vec-sphereC (as implemented by TorchSpatial), a spherical location encoder designed to preserve geodesic distances directly on the spherical surface across multiple spatial scales without map projection distortion.

Following results are reported as R@100 (%), mean ± std over three seeds:

Cityω\omega Range2D RFF seen3D RFF seensphereC seen2D RFF unseen3D RFF unseensphereC unseen
New York(0.05, 5)13.8 ± 0.912.7 ± 0.213.0 ± 0.316.2 ± 0.915.8 ± 0.415.9 ± 0.4
New York(0.1, 10)14.2 ± 1.413.0 ± 0.213.7 ± 0.315.8 ± 0.315.7 ± 0.416.0 ± 0.1
New York(0.5, 50)18.8 ± 1.414.6 ± 0.919.2 ± 0.714.6 ± 0.614.8 ± 0.615.2 ± 0.6
New York(5, 600)21.8 ± 2.022.5 ± 0.423.1 ± 0.18.6 ± 0.811.5 ± 0.89.9 ± 0.2
Istanbul(0.05, 5)38.8 ± 0.831.0 ± 2.435.0 ± 1.219.5 ± 0.118.7 ± 1.919.0 ± 0.2
Istanbul(0.1, 10)39.4 ± 2.436.9 ± 1.240.9 ± 0.519.6 ± 0.719.8 ± 0.419.4 ± 0.1
Istanbul(0.5, 50)47.4 ± 1.541.7 ± 0.445.4 ± 1.718.6 ± 0.719.7 ± 0.719.2 ± 0.6
Istanbul(5, 600)53.4 ± 0.851.1 ± 0.653.6 ± 0.214.6 ± 1.115.9 ± 1.013.6 ± 0.4

Across both cities and every frequency range, unseen recall under the planar treatment is within 1 percentage point of both geodesic-aware treatments. Where the treatments differ, the planar input is slightly better. We conclude that at the metropolitan scale, where the cos(latitude) anisotropy is effectively constant, using standardized lat/lon coordinates as input is a reasonable and practical choice.

W2 (novelty limited to the embedding). Our contribution is deliberately empirical: a controlled characterization of how the spatial scale of continuous location features governs the seen/unseen generalization tradeoff, with practical selection guidance. The abstract is written with this in mind.

W3 (benefit over the heuristic prior is weak). We agree with the reviewer that, on average, the benefit of the learned location encoder over the Gravity+history heuristic is modest. However, as shown in Appendix C, Figure 7, the RFF encoder's gain over Gravity+history is concentrated in the hardest prediction scenarios, e.g. where a user travels to an unpopular location, or more than 15 km from their last location. These scenarios are precisely where the learned embedding provides the most value.

W4 (R@100 too permissive; smaller and ranking-sensitive metrics). We will add the metrics the reviewer has suggested to our next revision. Below are the numbers for R@5, R@10, R@20, R@100, MRR and NDCG@10. Here, we report model-only scores (without any heuristics added):

EncoderCityGroupR@5R@10R@20R@100MRR×100NDCG@10×100
RFFNew Yorkunseen1.25 ± 0.152.45 ± 0.084.54 ± 0.2316.21 ± 0.851.21 ± 0.101.10 ± 0.08
Space2Vec-hexNew Yorkunseen1.23 ± 0.132.33 ± 0.234.27 ± 0.3015.91 ± 0.661.15 ± 0.071.04 ± 0.10
Space2Vec-gridNew Yorkunseen1.08 ± 0.202.17 ± 0.354.14 ± 0.3215.23 ± 1.131.08 ± 0.080.96 ± 0.15
RFFNew Yorkseen1.19 ± 0.192.24 ± 0.374.21 ± 0.4813.84 ± 0.921.12 ± 0.151.03 ± 0.19
Space2Vec-hexNew Yorkseen1.06 ± 0.351.94 ± 0.483.84 ± 0.7513.63 ± 1.181.07 ± 0.250.92 ± 0.29
Space2Vec-gridNew Yorkseen1.30 ± 0.152.42 ± 0.284.44 ± 0.2614.38 ± 0.901.30 ± 0.181.20 ± 0.19
RFFNew Yorkall1.21 ± 0.082.31 ± 0.254.32 ± 0.3814.64 ± 0.451.15 ± 0.071.05 ± 0.10
Space2Vec-hexNew Yorkall1.12 ± 0.252.07 ± 0.303.99 ± 0.5914.40 ± 1.001.10 ± 0.190.96 ± 0.20
Space2Vec-gridNew Yorkall1.22 ± 0.032.34 ± 0.074.34 ± 0.1314.67 ± 0.811.23 ± 0.091.12 ± 0.08
RFFIstanbulunseen1.26 ± 0.152.95 ± 0.455.59 ± 0.8319.47 ± 0.101.27 ± 0.091.17 ± 0.19
Space2Vec-hexIstanbulunseen1.18 ± 0.382.54 ± 0.335.23 ± 0.1320.23 ± 0.171.23 ± 0.121.03 ± 0.19
Space2Vec-gridIstanbulunseen1.14 ± 0.122.36 ± 0.195.32 ± 0.6418.95 ± 0.291.24 ± 0.061.01 ± 0.06
RFFIstanbulseen10.14 ± 1.2314.21 ± 0.8519.75 ± 0.5638.83 ± 0.757.67 ± 0.858.46 ± 0.89
Space2Vec-hexIstanbulseen9.66 ± 0.9813.56 ± 0.9918.70 ± 1.3636.85 ± 1.487.69 ± 0.698.36 ± 0.76
Space2Vec-gridIstanbulseen8.37 ± 1.2211.74 ± 1.1816.27 ± 1.5434.26 ± 2.016.81 ± 0.737.29 ± 0.84
RFFIstanbulall7.87 ± 0.9111.33 ± 0.6016.13 ± 0.5733.88 ± 0.546.03 ± 0.616.60 ± 0.63
Space2Vec-hexIstanbulall7.49 ± 0.6410.74 ± 0.6915.25 ± 0.9932.60 ± 1.106.04 ± 0.486.48 ± 0.52
Space2Vec-gridIstanbulall6.52 ± 0.889.34 ± 0.9213.46 ± 1.2330.34 ± 1.555.39 ± 0.555.69 ± 0.64

Here all combines seen and unseen targets at their actual test-set proportions.

W5 (code release). An anonymized release of the core implementation is now available: [ANONYMIZED-REPO-LINK].

W6 (computational efficiency and deployment cost). Every continuous encoder has 8.7–8.9M parameters (RFF 8.92M, Space2Vec-grid 8.92M, Space2Vec-hex 8.66M) independent of the number of POIs, versus a per-POI embedding table of 10.2M parameters on the New York dataset and 17.0M parameters on the Istanbul dataset.

Inference cost: Embedding the full New York candidate set (39,654 coordinates) takes ~0.6 s on a single laptop CPU core, and a single coordinate takes 0.4 ms.

Q1 (one embedding per location; POI semantics). We agree that a coordinate embedding cannot distinguish two POIs at the same location. Our paper's scope is intended to be limited to comparing geo-spatial encodings without introducing semantic information. We expect the fusion of semantic features such as the categories of POIs to be orthogonal to our work. However, these fusion methods remain good avenues for future work.

Q2 (does dropping 20% of POIs reflect naturally emerging POIs?). We agree with the reviewer that this is a fair concern. One way to address this concern would be to split the trajectories chronologically and evaluate on POIs that only appear in the test set. We considered this early on in the project, but noticed that, when split this way, the novel POIs tend to be rare POIs (in the tail of the popularity distribution) and are unevenly distributed geographically. Thus we found the H3-stratified sampling method to yield more reliable comparisons across methods.

We would like to thank the reviewer for their feedback, especially with regards to the treatment of latitude and longitude inputs, which helped us improve our paper substantially.

As the discussion period closes, please let us know if our reply sufficiently addressed your concerns, and if so, we would be grateful if you would consider updating your assessment. Thank you!

Thank you for your reply. Most of my questions are solved. I will raise my score to 3.

Reviewer_ntaMOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies generalization to locations absent from training in next point-of-interest prediction. It combines a continuous latitude–longitude encoder based on random Fourier features (RFF) with a causal transformer trained through contrastive learning. An extended training scheme incorporates multi-step supervision and heuristic-residual learning using distance, popularity, and visit history. The experiments compare location encoders across eight cities in Massive-STEPS and examine the tradeoff between seen and unseen POI recall under location holdouts in six cities. Lowering the RFF frequency range increases average unseen Recall@100 from 36.4% to 42.9%, while reducing seen recall from 63.8% to 57.5%. Additional experiments examine this tradeoff across training seeds and first-view predictions, with a real-estate location prediction study providing supplementary evidence.

Critical strengths

S1. The frequency–generalization tradeoff has substantial empirical support. Table 3 shows that reducing the RFF frequency range from (5, 600) to (0.1, 10) improves unseen recall in all six cities, while reducing average seen recall. Table 7 in Appendix C.1 reproduces the direction of this tradeoff across six cities and three training seeds using model scores alone, showing that it does not depend entirely on heuristic ensembling. Further reducing the frequency range does not consistently help. These results provide concrete guidance on choosing the spatial resolution of location representations and are the main basis for my positive assessment.

S2. The comparison of location representations is useful for model design. Within a common sequence-modeling framework, the paper compares per-POI, administrative-region, H3, and continuous coordinate encoders. In Table 1, RFF achieves average POI Recall@100 of 62.5%, compared with 60.7% for per-POI embeddings and 59.3% for Gravity+history. Table 2 also examines performance at different H3 resolutions. These experiments support the usefulness of continuous representations in this framework and show how their benefit varies with evaluation granularity. The approach requires only coordinates and trajectories, which makes it applicable when additional POI attributes are unavailable.

Critical weaknesses

W1. The additional benefit for unseen-location prediction remains limited, and an advantage over existing continuous encoders is not firmly established. In Table 3, the lower-frequency RFF configuration achieves 42.9% average unseen recall versus 42.2% for Gravity+history, a gain of 0.7 percentage points accompanied by lower seen recall. In the stricter first-view evaluation in Appendix A, standalone coarse RFF remains below Spatial, and validation-selected combinations with Spatial do not establish a consistent improvement. Removing inference-time popularity helps in Istanbul but does not produce a comparable improvement in New York. Furthermore, RFF exceeds Space2Vec-hex by only 0.2 percentage points in average Recall@100 in Table 1, while only RFF receives the full frequency sweep. The authors acknowledge several of these limitations. Nevertheless, they make the contribution primarily an empirical analysis of spatial scale, with more limited evidence for additional gains in recommending new locations.

W2. The real-estate experiment does not yet provide independent validation under a realistic prediction protocol. Section 4.2 reports validation-set results and constructs candidate locations using the true transaction city. Appendix B also states that the popularity KDE uses training locations belonging to that city. The common ensemble weight is selected to maximize trainable RFF's validation MRR. This provides an exploratory comparison under a particular candidate construction scheme, but does not establish whether the gains persist when the eventual transaction city is unknown and final evaluation is separate from model selection. The authors explicitly acknowledge this scope in the Limitations section, so I give this experiment limited weight as evidence for national-scale deployment.

Questions for authors

Q1. Clarification: Were the spatial-scale parameters of Space2Vec-grid and Space2Vec-hex given a search budget comparable to RFF? If such results are available, how do the seen/unseen tradeoff in Table 3 and first-view unseen performance compare with tuned Space2Vec at similar model capacity? This would help distinguish the benefit of RFF itself from the benefit of spatial-scale selection.

Q2. Clarification: Does the real-estate dataset include a test set independent of hyperparameter selection and candidate sets generated without using the eventual transaction city? If this evaluation is available, do the RFF gains persist? Otherwise, please clarify that the reported experiment evaluates ranking among candidates constructed using the true transaction city, and specify which city information would be available before prediction in the intended application.

2 comments

We thank the reviewer for the thorough review and for highlighting the empirical support of the frequency–generalization tradeoff as the basis of their positive assessment.

W1 and Q1 (advantage over existing continuous encoders; matched search budget for Space2Vec). We agree this was a key missing comparison, and we now run exactly the experiment suggested by the reviewer: we vary the ω\omega frequency parameter in Space2Vec-grid and Space2Vec-hex over the same four wavelength ranges as RFF, on the New York and Istanbul datasets under the Table 3 protocol. These results are below:

Encoderω\omega RangeNew York SeenNew York UnseenIstanbul SeenIstanbul Unseen
Space2Vec-grid(0.05, 5)13.5 ± 0.514.7 ± 1.234.3 ± 2.018.9 ± 0.3
Space2Vec-grid(0.1, 10)14.4 ± 0.915.2 ± 1.134.5 ± 3.019.3 ± 0.6
Space2Vec-grid(0.5, 50)18.4 ± 1.115.9 ± 0.244.8 ± 1.619.4 ± 0.6
Space2Vec-grid(5, 600)23.0 ± 1.09.9 ± 0.452.8 ± 0.812.2 ± 0.2
Space2Vec-hex(0.05, 5)13.6 ± 1.215.9 ± 0.736.9 ± 1.520.2 ± 0.2
Space2Vec-hex(0.1, 10)14.5 ± 1.015.8 ± 0.138.2 ± 4.919.6 ± 1.2
Space2Vec-hex(0.5, 50)18.9 ± 0.815.4 ± 0.244.5 ± 1.519.3 ± 0.8
Space2Vec-hex(5, 600)22.3 ± 0.59.6 ± 0.552.8 ± 0.912.9 ± 0.9
RFF(0.05, 5)13.8 ± 0.916.2 ± 0.938.8 ± 0.819.5 ± 0.1
RFF(0.1, 10)14.2 ± 1.415.8 ± 0.339.4 ± 2.419.6 ± 0.7
RFF(0.5, 50)18.8 ± 1.414.6 ± 0.647.4 ± 1.518.6 ± 0.7
RFF(5, 600)21.8 ± 2.08.6 ± 0.853.4 ± 0.814.6 ± 1.1

To answer Q1 directly based on these results, the same seen/unseen tradeoff replicates for Space2Vec as ω\omega varies. The performance difference in the methods can be primarily attributable to the spatial scale selection, which is the primary claim of our submission.

W2 and Q2 (real-estate protocol: candidates conditioned on the true transaction city). The reviewer's reading is correct: the reported experiment evaluates ranking among candidates constructed using the true transaction city, with ensemble weights selected on validation data. Unfortunately, we cannot run this experiment without knowledge of the eventual transaction city during the rebuttal period. The intention of this experiment was to replicate our real-world production setup as much as possible. Since transaction city prediction is already a solved problem and a mature part of the production system, we believe that proving the benefit of continuous location embeddings on top of the existing system is more beneficial to the industry.

Thank you again for the positive review, and we hope these results address your concerns.

We would like to thank the reviewer for their feedback, which helped us improve our paper substantially.

As the discussion period closes, please let us know if our reply sufficiently addressed your concerns, and if so, we would be grateful if you would consider updating your assessment. Thank you!

Reviewer_EqfMOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a framework for next-location prediction that aims to generalize to previously unseen points of interest (POIs). The proposed approach employs random Fourier features (RFF) followed by an MLP to encode latitude–longitude coordinates into continuous representations that capture spatial relationships between locations. A transformer then processes the user's historical check-in sequence to predict the embedding of the next location. Unlike conventional methods that learn separate embeddings for individual POIs, this framework generates location representations directly from coordinates, enabling predictions for locations not encountered during training.

Overall, the paper provides an interesting empirical analysis of how spatial representation granularity affects generalization in next-location prediction. However, given that several existing approaches also employ continuous coordinate-based representations, a clearer justification of the methodological advantages of RFF over alternative location encoders would help strengthen the paper's contribution.

Critical strengths
  1. Parameter efficiency and practical insights:
  • By representing locations directly through latitude–longitude coordinates, the proposed framework eliminates the need to learn separate embeddings for individual POIs. Consequently, the number of location encoder parameters remains independent of the size of the POI catalog, making the approach potentially scalable to settings with continuously expanding locations.
  • Furthermore, the empirical analysis of RFF frequency scales provides useful practical insights into how spatial representation granularity can be adjusted to balance prediction accuracy on seen and unseen locations.
Critical weaknesses
  1. Limited methodological novelty:
  • Although the paper investigates the use of RFF for next-location prediction, several existing approaches also employ continuous representations derived from latitude–longitude coordinates. The methodological distinction between RFF and these alternative location encoders is not sufficiently clear.
  • While the paper provides a useful empirical analysis of RFF frequency settings, a stronger justification of its unique advantages over existing continuous location encoding methods would help establish the novelty and significance of the proposed approach.
  1. Marginal improvements over existing methods:
  • The performance gains over existing continuous location encoders are relatively modest. For example, the average POI-level improves by very small points compared with Space2Vec-hex.
  • Given the small performance differences, it remains unclear whether RFF offers a meaningful advantage over existing approaches. More rigorous comparisons, including statistical significance analyses, would help substantiate the claimed benefits.
  1. Trade-off between seen and unseen POI performance:
  • The experiments demonstrate that reducing RFF frequency scales improves recall on unseen POIs but simultaneously decreases recall on previously observed POIs.
  • In practical recommendation scenarios, both seen and unseen locations contribute to overall recommendation quality, and the relative importance of each depends on their prevalence in actual user behavior. Therefore, improvements in unseen-POI recall alone may not be sufficient to establish the overall effectiveness of the proposed approach.
Questions for authors
  1. What are the key methodological advantages of RFF compared with existing continuous location encoders? Could the authors clarify whether the baseline methods were sufficiently optimized? In particular, would the relative performance advantages of RFF persist if competing methods were tuned with comparable effort and appropriately selected hyperparameters?

  2. The average improvement of RFF over Space2Vec-hex is relatively small. Beyond statistical significance, is there additional evidence demonstrating that RFF provides meaningful advantages over Space2Vec? For example, Figure 6 provides qualitative comparisons between RFF and discrete location encoders, but does not include Space2Vec. A similar visualization comparing RFF with Space2Vec would help clarify whether RFF produces meaningfully different or more informative spatial preference distributions.

  3. Could the authors further analyze how improvements in unseen-POI recall translate into overall recommendation performance? Specifically, when both seen and unseen POIs are evaluated together under their actual proportions in the test data, do the gains in unseen recall outweigh the corresponding reductions in seen recall?

3 comments

We thank the reviewer for recognizing the parameter-efficiency of the approach and the practical value of the frequency-scale analysis. We appreciate the critical feedback and address each of the weaknesses and questions below.

W1 (methodological distinction from existing continuous encoders) and Q1 (were baselines sufficiently optimized?). We agree with the reviewer that the continuous encoder baselines should be subject to the same rigorous hyperparameter tuning as RFF. The hyperparameter that caused the largest difference in recall results was the frequency parameter ω\omega. Therefore, for the rebuttal we vary ω\omega for the two variations of Space2Vec to show that they exhibit exactly the same frequency–generalization tradeoff characteristic as RFF. Space2Vec-grid and Space2Vec-hex were swept over the same ω\omega ranges as RFF — (0.05, 5), (0.1, 10), (0.5, 50), (5, 600).

The results here follow similar settings as Table 3 in the main paper, measuring R@100 for both seen and unseen POIs:

Encoderω\omega RangeNew York SeenNew York UnseenIstanbul SeenIstanbul Unseen
Space2Vec-grid(0.05, 5)13.5 ± 0.514.7 ± 1.234.3 ± 2.018.9 ± 0.3
Space2Vec-grid(0.1, 10)14.4 ± 0.915.2 ± 1.134.5 ± 3.019.3 ± 0.6
Space2Vec-grid(0.5, 50)18.4 ± 1.115.9 ± 0.244.8 ± 1.619.4 ± 0.6
Space2Vec-grid(5, 600)23.0 ± 1.09.9 ± 0.452.8 ± 0.812.2 ± 0.2
Space2Vec-hex(0.05, 5)13.6 ± 1.215.9 ± 0.736.9 ± 1.520.2 ± 0.2
Space2Vec-hex(0.1, 10)14.5 ± 1.015.8 ± 0.138.2 ± 4.919.6 ± 1.2
Space2Vec-hex(0.5, 50)18.9 ± 0.815.4 ± 0.244.5 ± 1.519.3 ± 0.8
Space2Vec-hex(5, 600)22.3 ± 0.59.6 ± 0.552.8 ± 0.912.9 ± 0.9
RFF(0.05, 5)13.8 ± 0.916.2 ± 0.938.8 ± 0.819.5 ± 0.1
RFF(0.1, 10)14.2 ± 1.415.8 ± 0.339.4 ± 2.419.6 ± 0.7
RFF(0.5, 50)18.8 ± 1.414.6 ± 0.647.4 ± 1.518.6 ± 0.7
RFF(5, 600)21.8 ± 2.08.6 ± 0.853.4 ± 0.814.6 ± 1.1

Tuned Space2Vec exhibits exactly the frequency–generalization tradeoff our submission characterizes for RFF: in every encoder and both city datasets, coarser ranges raise unseen recall and lower seen recall, and the finest range collapses unseen recall.

W2 (marginal improvements) We would like to point to evidence already in the submitted paper: Table 7 reports mean ± std over three training seeds for the key frequency-tradeoff result, showing the results directionally hold in every city and seed. For the new Space2Vec results above, we report mean ± std across seeds. We agree that the 0.2-point Table 1 average gap over Space2Vec-hex alone does not establish superiority, and the paper already acknowledges this as a limitation. We reiterate that the main claim the submission defends is that tuning the frequency parameter in the location encoder traverses a tradeoff between seen and unseen recall.

W3 and Q3 (does unseen gain outweigh seen loss under realistic proportions?). The real-world test-set proportions between seen and unseen POIs are: 66.3%/33.7% in New York (10,745/5,457 events) and 74.4%/25.6% in Istanbul (19,891/6,843). The all group in the table below ranks them jointly.

CityRFF rangeSeenUnseenCombined (all)
New York(0.05, 5)13.8 ± 0.916.2 ± 0.914.6 ± 0.5
New York(0.1, 10)14.2 ± 1.415.8 ± 0.314.7 ± 0.9
New York(0.5, 50)18.8 ± 1.414.6 ± 0.617.4 ± 1.1
New York(5, 600)21.8 ± 2.08.6 ± 0.817.3 ± 1.4
Istanbul(0.05, 5)38.8 ± 0.819.5 ± 0.133.9 ± 0.5
Istanbul(0.1, 10)39.4 ± 2.419.6 ± 0.734.3 ± 2.0
Istanbul(0.5, 50)47.4 ± 1.518.6 ± 0.740.0 ± 1.0
Istanbul(5, 600)53.4 ± 0.814.6 ± 1.143.5 ± 0.6

The table shows a similar tradeoff: coarser scales favor unseen recall, whereas finer scales favor combined recall when seen targets dominate. We do not claim the lower-frequency setting is uniformly better.

Importantly, also note that Table 1 of the submitted paper already evaluates using the actual seen/unseen proportions on a group of randomly withheld users.

We hope that these results address the concerns behind the score.

Thank you for the authors’ response. The additional analysis and experiments provided by the authors have addressed most of my concerns. However, I still find the motivation for the proposed approach to be marginal. Therefore, I am willing to raise my score to 2.

We would like to thank the reviewer for their insightful feedback and engagement during the discussion phase. We really appreciate your service!

Comments on the paper

1 comment
AuthorsOpenReview ↗

Overall Rebuttal Comment

We thank all reviewers for their thoughtful reviews. We are glad the reviewers found the frequency–generalization tradeoff analysis to have "substantial empirical support" and to "provide concrete guidance on choosing the spatial resolution of location representations" (Reviewer ntaM), the parameter-efficiency and granularity analysis to offer "useful practical insights" (Reviewer EqfM), and the problem of recommending newly introduced POIs to be "a practical problem" (Reviewer rsmp).

One common point was raised across all three reviews, we address that briefly here before responding to each reviewer individually.

Were competing continuous encoders (such as Space2Vec) tuned as thoroughly as RFF? (Reviewer EqfM-W1/Q1, Reviewer ntaM-W1/Q1, Reviewer rsmp-W3). For the rebuttal we ran the same hyperparameter search over the range of frequencies ω\omega for Space2Vec-grid and Space2Vec-hex as we did for RFF. The results show that varying the frequency parameter traverses the same seen-unseen recall tradeoff curve as RFF. This is the central claim in our abstract, and the rebuttal shows that it generalizes to Space2Vec. We provide numerical results within the individual reviewer responses.

We have also released an anonymized version of our full codebase: [ANONYMIZED-REPO-LINK]

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.