acceptodds
Under review as a conference paper at ICLR 2027

STCD-Hawkes: Spatio-Temporal Causal Debiasing with Hawkes Process for Next POI Recommendation

Abstract

Next-point-of-interest recommendation is essential for location-based services but remains challenging because of the complex spatio-temporal characteristics of user mobility. Existing methods often treat temporal and spatial information as static features or rely on simple fusion strategies, overlooking both the distributional heterogeneity between temporal triggers and spatial choices and the causal entanglement in which contextual constraints, such as distance, obscure users’ intrinsic preferences. To address these limitations, we propose STCD-Hawkes, a unified probabilistic framework for spatio-temporal causal debiasing with Hawkes processes. The framework decomposes next-POI recommendation into two coupled generative processes. To model when a user moves, it employs a time-aware Hawkes process with a rotation-based adaptation mechanism to capture the continuous evolution of temporal intensity. To predict where the user goes, it introduces a score-based latent diffusion model that generates optimal spatial representations from noise under the condition of the learned temporal intensity. To mitigate confounding effects, it further incorporates a causal disentanglement module based on a disentangled causal mixture-of-experts architecture with orthogonal regularization, explicitly separating invariant user interests from dynamic contextual sensitivities. Extensive experiments on three real-world datasets demonstrate that STCD-Hawkes significantly outperforms state-of-the-art baselines while providing interpretable insights into the causal factors underlying user mobility.

Primary area:
learning on time series and dynamical systems
Keywords:
Next POI Recommendation, Spatio‑Temporal Modeling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_zWGUOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes the unified probabilistic framework STCD‑Hawkes. It decomposes POI recommendation into two coupled generative processes: a time‑aware Hawkes process with rotation‑based adaptation mechanism is used to model continuous‑time intensity for "when to move"; a score‑based latent diffusion model generates spatial representations conditioned on the learned temporal intensity for "where to go". Furthermore, a disentangled causal Mixture‑of‑Experts (MoE) module with orthogonal regularization separates users’ intrinsic preferences from dynamic contextual sensitivities to achieve causal debiasing.

Critical strengths

It decomposes POI recommendation into two coupled sub‑processes: temporal intensity modeling and conditional spatial generation. The temporal module adopts a Hawkes process with rotation mechanism to model continuous and non‑uniform triggering intensity of user activities. The spatial module leverages score‑conditional diffusion to generate spatial representations instead of simple classification over fixed POI candidates. It explicitly addresses the distribution mismatch between temporal and spatial signals, and adapts to user trajectories of varying lengths.

Critical weaknesses

W1. The core innovations are limited, and the overall design resembles a composition of existing methods.

W2. Orthogonal regularization only imposes constraints in the representation space and achieves representational disentanglement rather than genuine causal intervention. Only counterfactual‑sensitivity heatmap analysis is provided, without intervention experiments (e.g., modifying contextual conditions and observing corresponding changes in recommendations). Thus, it cannot rigorously prove that proximity‑induced confounding effects have been fully eliminated.

W3. The selected baselines lack comparisons against existing diffusion‑based POI recommendation methods, such as Diff‑POI, Diff‑DGMN, UB‑Diff, DiffuRec, and PreferDiff.

Questions for authors

Q1. Spatial generation relies on the score‑based SDE diffusion model which requires iterative sampling. It incurs higher training overhead compared with pure Transformer‑based and graph‑based baselines. What is the actual training time of your model?

Q2. For real‑world online POI recommendation services, how do you balance the generation quality from diffusion and inference efficiency?

Q3. Orthogonal regularization in this work only yields approximate disentanglement within the feature space. Have you considered conducting real causal intervention experiments, for instance artificially modifying distance‑related context and checking whether user intrinsic preferences can still be preserved in recommendations, to further validate the effectiveness of your causal debiasing?

2 comments

W1. Thank you for your evaluation of our contribution. STCD-Hawkes is built on a principled coupling of continuous-time activity modeling, spatial score learning, and structured expert routing. Rather than a naive composition, these components interact through a common learned representation and are jointly optimized via Eq. (19). Table 3 evaluates their individual predictive contributions within this unified architecture.

W2 and Q3. We appreciate your insights on causal motivation. We agree with the reviewer that orthogonal regularization and counterfactual sensitivity analyses serve as structural inductive biases and local context-sensitivity diagnostics, rather than exhaustive causal intervention experiments. In the revised manuscript, we will moderate strong causal phrasing and explicitly clarify the scope of these diagnostics, keeping them distinct from module-level predictive comparisons.

W3. Thank you for the pointer to recent diffusion-based models. While these methods offer valuable perspectives, STCD-Hawkes is uniquely distinguished by its explicit event-intensity formulation combined with time-adapted, expert-based spatial conditioning. We will expand the Related Work section to thoroughly discuss and contextualize these baselines.

Q1 and Q2. Thank you for your questions regarding computational cost. Thanks to the closed-form VP-SDE transition kernel (Eq. 10), training does not require simulating intermediate forward diffusion states. For inference, we control computational overhead by setting the diffusion steps to T=10T=10 (Appendix D.3), striking an optimal balance between generative fidelity and online efficiency. We will include a clear discussion on training and inference complexity in the revision.

Reviewer_zWGUOpenReview ↗

Reply to Authors

Thank you for your rebuttal. My concerns still remain, as the core innovations are weak. Secondly, the author's response did not effectively address my Questions. I therefore maintain my score of 1.

Reviewer_9jxmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes STCD-Hawkes for next-POI recommendation. It combines a Transformer trajectory encoder, a temporal intensity model with a rotation mechanism, and conditional diffusion in the spatial embedding space. A mixture-of-experts module is introduced to separate user preferences from contextual influences. The method is evaluated on NYC, TKY, and CA, with comparisons against existing methods, component ablations, and analyses of trajectory length.

Critical strengths
  1. The paper is well motivated by the irregular timing of check-ins and the dependence of destination choice on the current context. The rotation mechanism in Eqs. (5)-(8) is described clearly and provides a way to adapt the historical representation as time passes.
  2. Table 1 reports improvements on all four metrics across three datasets. The gains on NYC and CA are substantial: Acc@1 rises from 25.69% to 31.10% and from 16.47% to 20.38%, respectively, compared with the strongest included baseline. The chronological split is appropriate for this task. The component ablations and results by trajectory length also help assess where the method is useful.
Critical weaknesses
  1. It is unclear how the spatial diffusion module is used for next-POI prediction. Eq. (20) scores candidates with the intrinsic and contextual representations, but contains no generated spatial representation or density. The diffusion loss may still help train the shared encoder. Please clarify whether reverse diffusion is run at test time and how its output affects the scores. The source of the target time in Fig. 1 also needs to be stated. These details are needed to interpret the reported results and the diffusion-step analysis in Appendix H.2.
  2. The causal interpretation needs stronger support. Section 2.4.2 states that orthogonal expert weights ensure statistically independent representations, which does not follow in general. Both representations also include the same shared-expert output in Eqs. (14)-(15). In Appendix F, Eq. (23) measures local sensitivity to input changes; it does not establish a causal intervention effect. The ablation in Table 3 removes the whole MoE module, so the improvement could also reflect additional model capacity. Please provide evidence that the learned components correspond to stable preferences and contextual effects, or limit the claims to improved prediction.
Questions for authors
  1. Please describe the test-time procedure, including how the target time in Fig. 1 is obtained and how the output of reverse diffusion enters Eq. (20). If the query time is known, were the baselines given the same information? If diffusion is used only during training, please explain how the reverse-sampling step count affects the results in Appendix H.2.
  2. Would the gains remain with the same MoE architecture but without the orthogonality constraint? This comparison would help separate the effect of regularization from that of model capacity. Please also explain what evidence supports interpreting the two outputs as intrinsic preferences and contextual effects.
  3. Please check the TKY results in Tables 1 and 3 and Figs. 5 and 6. The tables report Acc@1 = 25.11%, whereas the curves at the stated default settings appear to give about 37%. Some plotted MRR values also appear lower than Acc@1, which is inconsistent with the stated metric definitions for the same set of predictions. Please provide the exact plotted values, explain any differences in settings, and state whether the main results need correction.
  4. The contribution relative to existing work needs clearer discussion. Chen et al. [1] already use the decomposition into temporal intensity and conditional spatial density. Diff-POI [2], which is cited but not included in Table 1, uses score-based diffusion to sample spatial preferences for POI ranking. Please explain what STCD-Hawkes adds to these approaches and compare with Diff-POI under the same evaluation protocol, or explain why it was not included.

References

[1] Ricky T. Q. Chen, Brandon Amos, and Maximilian Nickel. Neural Spatio-Temporal Point Processes. ICLR, 2021. https://arxiv.org/abs/2011.04583

[2] Yifang Qin, Hongjun Wu, Wei Ju, Xiao Luo, and Ming Zhang. A Diffusion Model for POI Recommendation. ACM Transactions on Information Systems, 42(2):1-27, 2024 (published online in 2023). https://doi.org/10.1145/3624475

1 comment

Q1. Thank you for your question regarding the spatial diffusion module. During training, the diffusion objective jointly shapes the latent representation space used by the ranking head (Eq. 20). At inference time, reverse diffusion generates expected spatial representations as geometric guidance vectors to complement semantic matching. The target time is naturally derived from the Hawkes intensity process λ(t∣Ht)\lambda(t\vert{}\mathcal{H}_t). We will explicitly clarify this end-to-end inference pipeline and timestamp flow in the revision.

Q2. We appreciate your attention to the expert representations. Our orthogonal regularization (Lorth\mathcal{L}_{orth}) goes beyond model capacity by structurally penalizing cross-group weight overlaps, preventing feature leakage and mitigating proximity confounding bias. This disentanglement is empirically supported by our Counterfactual Sensitivity Analysis (Appendix F), showing that contextual outputs dynamically track local triggers while invariant preferences remain stable. We will refine the terminology in the text to "structural causal disentanglement via orthogonal subspace routing."

Q3. Thank you for pointing out this discrepancy. We sincerely apologize for the confusion caused by a plotting and numerical measurement error in Figures 5 and 6. The metrics in Tables 1 and 3 (Acc@1=25.11%,MRR=35.20%Acc@1=25.11\%, MRR=35.20\% on TKY) are the accurate, verified benchmark results. Due to an oversight, Figures 5 and 6 incorrectly displayed the baseline values. We will carefully correct and update these figures in the revision to ensure numerical consistency with the main evaluation tables.

Q4. Thank you for the suggestion to clarify prior work. Unlike Chen et al. (general event streams) or Diff-POI (graph-based spatial diffusion), STCD-Hawkes uniquely unifies continuous-time Hawkes intensity ("when"), score-based spatial diffusion ("where"), and causal MoE debiasing ("why") into an end-to-end probabilistic framework. We will expand the Related Work and Discussion sections to clearly contrast these architectural distinctions and baseline scopes.

Reviewer_2z3wOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes STCD‑Hawkes, a unified probabilistic framework for next Point‑of‑Interest (POI) recommendation that integrates continuous‑time Hawkes process modeling, score‑based conditional spatial diffusion, and causal‑flavored representation disentanglement. To model user mobility, the authors decompose joint intensity into temporal intensity and conditional spatial density. A rotation‑adapted Time‑Aware Hawkes Process captures continuous temporal dynamics, a score‑based latent diffusion module generates spatial representations conditioned on temporal states, and a disentangled Mixture‑of‑Experts module aims to separate intrinsic user preferences from context‑driven situational effects. Extensive experiments are conducted on three real‑world POI check‑in datasets, showing overall improvements against multiple state‑of‑the‑art baselines.

Critical strengths

The decomposition of user mobility into temporal intensity and conditional spatial distribution provides a coherent perspective for spatio‑temporal POI recommendation. Integrating temporal modeling, spatial generation, and disentanglement of intrinsic preferences from contextual influences into a unified framework is a potentially useful design.

Critical weaknesses

The paper’s causal‑oriented interpretation is insufficiently justified. Applying orthogonal regularization on expert weight matrices cannot guarantee statistical independence between the learned intrinsic and contextual representations.

Questions for authors

What assumptions or additional empirical evidence support the causal interpretation of the disentangled intrinsic and contextual representations? If rigorous causal identification cannot be achieved, please moderate causal‑related wording across the manuscript.

1 comment

Modeling premise and expert roles. Thank you for raising this point. We agree that strict statistical independence cannot be universally guaranteed via orthogonal regularization alone in observational trajectory data. Rather than claiming absolute causal identification, our design uses the intrinsic, contextual, and shared expert groups as a structural inductive bias to separate stable preferences from dynamic contextual triggers, while retaining overlapping components via shared experts.

Clarification on causal terminology. We sincerely appreciate the reviewer's suggestion. In the revised manuscript, we will moderate strong causal-related claims and reframe the framework's objective more precisely—describing it as an effective inductive bias for representation disentanglement rather than a formal proof of causal independence.

Empirical evidence and diagnostics. Thank you for your feedback on the empirical evaluation. We agree that the counterfactual sensitivity analysis in Appendix F serves as a local context-sensitivity diagnostic rather than a standalone causal intervention experiment. We will explicitly clarify this scope in the revised text, ensuring that our empirical diagnostics and predictive performance gains (Table 3) are interpreted with appropriate rigor.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.