- 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
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):
| Encoder | City | Group | R@5 | R@10 | R@20 | R@100 | MRR×100 | NDCG@10×100 |
|---|---|---|---|---|---|---|---|---|
| RFF | New York | unseen | 1.25 ± 0.15 | 2.45 ± 0.08 | 4.54 ± 0.23 | 16.21 ± 0.85 | 1.21 ± 0.10 | 1.10 ± 0.08 |
| Space2Vec-hex | New York | unseen | 1.23 ± 0.13 | 2.33 ± 0.23 | 4.27 ± 0.30 | 15.91 ± 0.66 | 1.15 ± 0.07 | 1.04 ± 0.10 |
| Space2Vec-grid | New York | unseen | 1.08 ± 0.20 | 2.17 ± 0.35 | 4.14 ± 0.32 | 15.23 ± 1.13 | 1.08 ± 0.08 | 0.96 ± 0.15 |
| RFF | New York | seen | 1.19 ± 0.19 | 2.24 ± 0.37 | 4.21 ± 0.48 | 13.84 ± 0.92 | 1.12 ± 0.15 | 1.03 ± 0.19 |
| Space2Vec-hex | New York | seen | 1.06 ± 0.35 | 1.94 ± 0.48 | 3.84 ± 0.75 | 13.63 ± 1.18 | 1.07 ± 0.25 | 0.92 ± 0.29 |
| Space2Vec-grid | New York | seen | 1.30 ± 0.15 | 2.42 ± 0.28 | 4.44 ± 0.26 | 14.38 ± 0.90 | 1.30 ± 0.18 | 1.20 ± 0.19 |
| RFF | New York | all | 1.21 ± 0.08 | 2.31 ± 0.25 | 4.32 ± 0.38 | 14.64 ± 0.45 | 1.15 ± 0.07 | 1.05 ± 0.10 |
| Space2Vec-hex | New York | all | 1.12 ± 0.25 | 2.07 ± 0.30 | 3.99 ± 0.59 | 14.40 ± 1.00 | 1.10 ± 0.19 | 0.96 ± 0.20 |
| Space2Vec-grid | New York | all | 1.22 ± 0.03 | 2.34 ± 0.07 | 4.34 ± 0.13 | 14.67 ± 0.81 | 1.23 ± 0.09 | 1.12 ± 0.08 |
| RFF | Istanbul | unseen | 1.26 ± 0.15 | 2.95 ± 0.45 | 5.59 ± 0.83 | 19.47 ± 0.10 | 1.27 ± 0.09 | 1.17 ± 0.19 |
| Space2Vec-hex | Istanbul | unseen | 1.18 ± 0.38 | 2.54 ± 0.33 | 5.23 ± 0.13 | 20.23 ± 0.17 | 1.23 ± 0.12 | 1.03 ± 0.19 |
| Space2Vec-grid | Istanbul | unseen | 1.14 ± 0.12 | 2.36 ± 0.19 | 5.32 ± 0.64 | 18.95 ± 0.29 | 1.24 ± 0.06 | 1.01 ± 0.06 |
| RFF | Istanbul | seen | 10.14 ± 1.23 | 14.21 ± 0.85 | 19.75 ± 0.56 | 38.83 ± 0.75 | 7.67 ± 0.85 | 8.46 ± 0.89 |
| Space2Vec-hex | Istanbul | seen | 9.66 ± 0.98 | 13.56 ± 0.99 | 18.70 ± 1.36 | 36.85 ± 1.48 | 7.69 ± 0.69 | 8.36 ± 0.76 |
| Space2Vec-grid | Istanbul | seen | 8.37 ± 1.22 | 11.74 ± 1.18 | 16.27 ± 1.54 | 34.26 ± 2.01 | 6.81 ± 0.73 | 7.29 ± 0.84 |
| RFF | Istanbul | all | 7.87 ± 0.91 | 11.33 ± 0.60 | 16.13 ± 0.57 | 33.88 ± 0.54 | 6.03 ± 0.61 | 6.60 ± 0.63 |
| Space2Vec-hex | Istanbul | all | 7.49 ± 0.64 | 10.74 ± 0.69 | 15.25 ± 0.99 | 32.60 ± 1.10 | 6.04 ± 0.48 | 6.48 ± 0.52 |
| Space2Vec-grid | Istanbul | all | 6.52 ± 0.88 | 9.34 ± 0.92 | 13.46 ± 1.23 | 30.34 ± 1.55 | 5.39 ± 0.55 | 5.69 ± 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.
Thank you
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.
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)) 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:
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.