acceptodds
Under review as a conference paper at ICLR 2027

IDDiff:Intent-aware Deterministic Diffusion Model for Next Location Prediction

Abstract

Next location prediction aims to infer users’ next destinations from their histori- cal check-in trajectories. Although diffusion models offer a promising direction due to powerful generative capabilities, existing diffusion-based approaches en- counter two critical challenges: (i) check-in trajectories naturally entangle visits driven by diverse intents. Existing methods rely on raw check-in trajectory repre- sentations as contextual conditions to guide diffusion models, which can lead to learned representations that are inconsistent with the user’s authentic intents, ulti- mately yielding suboptimal performance; (ii) our experiment reveals a limitation when applying standard Gaussian noise to POI recommendation. The injected randomness may disrupt personalized preference structures, limiting their abil- ity to model users’ behaviors. To address these challenges, we propose IDDiff, an Intent-aware Deterministic Diffusion framework for next location prediction. Specifically, the Sequence-aware Trajectory Representation Module constructs a user’s global trajectory-level intent representation. It then divides the trajectory into multiple local subtrajectories and clusters semantically similar sub-trajectory representations to construct trajectory intent prototypes, which serve as contextual conditions to guide the diffusion model. Furthermore, we develop a Intent Refine- ment Module to address the structural disruption and reconstruction uncertainty caused by stochastic noise. Rather than injecting random Gaussian perturbations, the module constructs a predefined and controllable preference transformation tra- jectory, thereby preserving personalized mobility preference structures. Extensive experiments on five real-world datasets demonstrate that IDDiff consistently out- performs state-of-the-art methods. The implementation code is publicly available at https://anonymous.4open.science/r/IDDiff-7B4F.

Primary area:
generative models
Keywords:
Next Location Prediction, Diffusion Models, Generative Recommendation, Intent Modeling

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_CwCbOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses the problem of next-location prediction and proposes the IDDiff model. By integrating location-distance-based graph convolutions, trajectory segmentation via DBSCAN, and a deterministic diffusion model, the approach first segments users' historical trajectories and clusters them to extract travel intentions. It then incorporates the geographical graph structure to derive location representations. Finally, a deterministic diffusion process is employed to iteratively generate a representation of the user's preferences for prediction. This framework aims to mitigate two key challenges: the presence of irrelevant intentions within a full trajectory, and the potential for random Gaussian noise to disrupt the underlying structure of user preferences. Through empirical evaluation, the authors demonstrate that deterministic diffusion outperforms stochastic diffusion. Furthermore, they compare the model against various baselines using real-world check-in datasets from five cities, and conduct ablation studies, hyperparameter analysis, and visualizations.

Critical strengths
  1. The observations in the paper are sound, the research problem is meaningful, and the structure is clear and easy to follow.

  2. Comprehensive experiments were conducted using real-world datasets from five cities.

Critical weaknesses
  1. The design intent of the proposed approach is not sufficiently clear. The logical connection between the specific design choices and the problems they aim to address remains obscure, and there is a lack of theoretical justification for their validity. For instance, while the authors propose segmenting and clustering trajectories to extract current intent and filter out irrelevant visits, the final intent selection process relies on matching against a global representation of the entire historical trajectory. Similarly, the deterministic diffusion module merely reconstructs an initial vector that has already been computed during the prediction phase. The overall methodology appears to be a collection of off-the-shelf components stitched together, failing to provide a principled explanation of why this specific configuration effectively solves the problems posed at the beginning of the paper.

  2. The "Related Work" section consists merely of a list of references, lacking any in-depth comparison to highlight the fundamental differences between the proposed approach and existing literature. Whether discussing recent methods for next-location prediction or diffusion models used in recommender systems, the authors provide only brief, individual summaries of each paper. They fail to critically analyze the shortcomings of existing designs or articulate the substantive, methodological breakthroughs offered by their own work.

  3. The experimental section is limited to a mere presentation of data, lacking both a rigorous analysis of the observed phenomena and plausible explanations for them; furthermore, the experimental setup is not clearly described. Throughout the main experiments, ablation studies, and hyperparameter analyses, the text largely repeats the numerical fluctuations shown in the charts and draws conclusions based on pre-existing, assumed motivations, without delving into the underlying reasons for these results.

Questions for authors
  1. It is recommended to include an explanation or theoretical analysis of the mechanisms behind each module within the methodology section, effectively bridging the research questions posed in the introduction with the specific model design. For instance, when extracting the current intent, one could incorporate information such as the current timestamp or data from recent visits, enabling the model to make decisions based on the current state.

  2. In the experiments section, please clarify the specific rules used for data cleaning, splitting, and sequencing. Furthermore, rather than merely reporting fluctuations in performance metrics, provide an in-depth analysis--grounded in the unique characteristics of data from different cities or specific prediction examples--of why certain modules or parameter settings lead to observed changes.

  3. Given that the model incorporates segmented clustering and multi-step diffusion iterations, it is advisable to provide details on the actual time required for training and inference, and to compare the computational overhead against that of the baseline models.

Reviewer_q3dROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes IDDiff for next POI recommendation. It uses sliding-window segmentation and DBSCAN clustering to construct intent prototypes, a POI distance graph to model geographical relationships, and a deterministic diffusion module that interpolates between two historical representations with intent-conditioned multi-step MLP updates. Experiments are conducted on five city-level datasets.

Critical strengths

S1. Exploration of deterministic diffusion for next POI recommendation. The paper explores replacing stochastic Gaussian perturbation with a deterministic interpolation path in diffusion-based next-POI recommendation, incorporating local mobility pattern guidance. This direction is motivated by a clear practical concern, namely that random perturbation may introduce instability, and it remains relatively underexplored for this task.

S2. A direct stochastic-versus-deterministic diffusion comparison. The paper includes a direct empirical comparison between stochastic diffusion and a deterministic variant. This comparison provides a useful reference point for discussing deterministic alternatives in next-POI recommendation and helps delineate the scope within which conclusions from such comparisons can be reasonably generalized.

Critical weaknesses

W1. The methodological claims could be better supported. For “Intent Prototype Construction,” a temporally contiguous sliding window does not align with the claimed temporal discontinuity of related visits, leaving unclear how intent is represented or effectively captured. The “Structure-Preserving” claim is under-explained, as endpoint semantics and selection basis, the structure preserved, correction of initial-representation inaccuracies, and the added value of multi-step iteration beyond general conditional transformation remain unclear. The claimed “limitation in diffusion models based on stochastic noise” is too broad, since one stochastic implementation does not show that stochastic diffusion approaches are generally ineffective, given possible differences across methods.

W2. The empirical evidence is limited. The ablation does not sufficiently validate effectiveness or isolate the internal contribution of the deterministic diffusion mechanism, and removing this core module yields only the smallest performance drop. The comparison also omits highly relevant methods such as MPDC[1] and UB-Diff[2], with MPDC especially comparable in setting and framework.

[1] Multi-Perspective Driven Expected Location Preferences for Next POI Recommendations, SIGIR ’26. [2] Bridging User Dynamic Preferences: A Unified Bridge-based Diffusion Model for Next POI Recommendation, TBD ’25.

Questions for authors

1. Do local windows and cluster prototypes actually identify intent?
A fixed-length sliding window is still temporally contiguous, so intents may be mixed within a window or split across windows. Short-segment encoding may reduce noise relative to long sequences, but this does not by itself constitute intent recognition. How is “intent” represented here, and how can the authors verify that the local windows and cluster prototypes capture meaningful behavioral intent rather than merely reducing noise?

2. What is the basis for DBSCAN parameter settings and clustering results?
The DBSCAN parameters are not reported. What are the settings and their basis? Are they selected on a validation set or set empirically? How sensitive are the intent clusters and representative prototypes to parameter changes, and are they stable?

3. What exactly does “Structure-Preserving” mean in the deterministic interpolation?
The paper claims that the deterministic process is intent structure-preserving. What specific structure is being preserved? Why can linear interpolation between L^u\hat{\mathbf{L}}_u and pˉu\bar{\mathbf{p}}_u preserve this structure? Why do these endpoint choices and this interpolation path realize this property? Why this path has specific semantic meaning rather than being merely a numerical change between two historical vectors is not sufficiently explained.

4. If L^u\hat{\mathbf{L}}_u may be inaccurate, why is it used as the direct reconstruction target?
The paper states that L^u\hat{\mathbf{L}}_u may be noisy or incomplete, yet uses it as the reconstruction target. What information or evidence does reconstructing this already known representation add? Since reconstructing a known target largely amounts to optimizing the representation under the cross-entropy loss, why not provide such information directly to L^u\hat{\mathbf{L}}_u, and what does the multi-step reconstruction actually contribute?

5. Independent value and empirical support of the multi-step deterministic process.
What does the multi-step deterministic process learn beyond directly feeding [Lˉu;Su][\bar{\mathbf{L}}_u;\mathbf{S}_u] into a network with comparable parameter count and computational cost? Is it mainly fusing two historical representations through a deeper nonlinear mapping and joint recommendation supervision? More generally, is the observed effect mainly due to the iterative steps, larger model capacity, the nonlinear expressive power of the MLP, the intent condition, or the recommendation objective?

6. After removing diffusion, the performance drop appears relatively limited.
Removing the entire diffusion module (“w/o Diff”) leads to a relatively small metric gap. Since deterministic diffusion is presented as a core design, this result needs further explanation. Variants in which the diffusion condition participates in the final logits show a larger effect. This appears to come not from the diffusion design itself but from the logits computation or recommendation supervision provided by window-segment features.

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

The paper proposes IDDiff, a framework for next location prediction from users’ historical check-in trajectories. It constructs intent prototypes by clustering encoded trajectory subsequences, incorporates geographical dependencies through a distance graph encoder, and refines user preferences through intent-conditioned deterministic diffusion. Evaluation on five Foursquare city datasets reports improvements over nine baselines, supported by component ablations, a comparison of stochastic and deterministic diffusion, and hyperparameter analyses.

Critical strengths
  1. Clustering local trajectory representations provides a plausible way to connect temporally separated but semantically related behaviors. This directly addresses intent mixing in user trajectories and gives the architecture a meaningful task-specific rationale. I think it provides a meaningful problem and settles it reasonably.

  2. I believe the experiments provided by the author are comprehensive. The evaluation covers five cities and several baseline families. The controlled diffusion comparison and component ablations provide useful initial evidence that the proposed design choices contribute to prediction performance.

Critical weaknesses

1.There appears to be an inconsistency in the results for IDDiff in Table 2. Under standard single-target evaluation and using the same rankings and averaging weights, for example, the LA results imply that NDCG@10−NDCG@5 should be at least approximately 0.00338, but Table 2 reports 0.0002. Please clarify the metric definitions and aggregation protocol, and verify these entries.

2.The w/o Diff ablation removes the entire refinement process. Although the paper compares diffusion variants and step counts, it lacks a capacity-matched one-step baseline with the same conditioning and reconstruction supervision. This leaves the specific advantage of the proposed multi-step mechanism unclear, weakening support for a central methodological contribution.

Questions for authors

1.The LA results in Table 2 appear inconsistent. Under standard single-target evaluation, the reported Acc@5 and Acc@10 imply a larger NDCG@10 − NDCG@5 difference than reported. Could the authors verify these results and clarify the evaluation protocol?

2.The reverse process in Eqs. (5)–(7) is not sufficiently clear. How is the time-step defined, and how does reconstructing the initial preference representation improve the final representation?

3.The benefit of multi-step deterministic restoration remains unclear. Could the authors provide additional experiments comparing multi-step and one-step restoration to validate the contribution of the iterative mechanism?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.