Where and When: Predict Next POI and Its Explicit Timestamp in Sequential Recommendation

Yuanbo Xu, Hongxu Shen, Yiheng Jiang, En WangIJCAI 2025 · DOI · OpenReview

TL;DR

This paper introduces TAPT, a novel multi-task learning framework for sequential Point-of-Interest (POI) recommendation that explicitly models both where a user will go next and when they will go. By treating timestamps as first-class citizens alongside POIs, TAPT aims to provide more timely and relevant recommendations, addressing a key limitation in existing sequential recommendation systems.

What Problem Does This Paper Address?

Sequential Point-of-Interest (POI) recommendation systems aim to predict the next location a user might visit based on their past check-in history. While these systems have advanced significantly, they often treat timestamps as secondary information or overlook them entirely. This leads to a disconnect between predicting what a user might do and when they might do it, resulting in recommendations that might be relevant in terms of location but not necessarily timely. For instance, recommending a restaurant for lunch at 10 PM is less useful than recommending it at noon. This paper tackles this gap by proposing a framework that simultaneously predicts the next POI and its associated timestamp, enabling more practical and timely recommendations for applications like personalized travel planning.

Methods

The core of the proposed solution is the Time-Aware POI Recommender with Timestamp Prediction (TAPT), a multi-task learning framework. TAPT operates in three main stages:

  1. Timestamp Decoupling and Encoding: Instead of treating timestamps as single values, TAPT decouples them into their constituent components: hour, minute, and second. These components are then encoded into high-dimensional embeddings. This allows the model to capture fine-grained temporal patterns and nuances in user behavior.
  2. Joint Embedding: The encoded POI embeddings and timestamp embeddings are concatenated to form a "joint embedding." This unified representation allows the model to learn the intricate correlations between where a user goes and when they go there.
  3. Multi-Task Backbone: The joint embeddings are fed into a standard sequential recommendation backbone (e.g., Transformer-based models). This backbone is augmented with two specialized modules:
  4. POI Recommender Module: This module predicts the next POI based on the learned joint embeddings.
  5. Timestamp Predictor Module: This module, built upon the same backbone, is specifically designed to predict the timestamp of the next interaction.

The framework is trained using a unified loss function that balances the POI recommendation loss and the timestamp prediction loss, ensuring both tasks are optimized simultaneously.

Datasets

The authors evaluated TAPT on three publicly available Location-Based Social Network (LBSN) datasets:

  • Foursquare (NYC and Tokyo): These datasets capture user check-ins in urban environments.
  • Gowalla: This dataset includes user check-ins with geographical information.

The datasets were preprocessed by filtering out inactive users and unpopular POIs, and then split into training, validation, and testing sets.

Key Insights and Ideas

  • Timestamps are Crucial for Timeliness: The paper strongly argues that simply recommending the next POI without considering when it will be visited leads to suboptimal and less practical recommendations.
  • Decoupling Timestamps Unlocks Granularity: Breaking down timestamps into hour, minute, and second allows for a much richer understanding of temporal user behavior than treating them as a single continuous value.
  • Joint Learning for Synergy: By jointly learning POI recommendation and timestamp prediction, TAPT leverages the interdependencies between location and time, leading to improved performance in both tasks.
  • Timestamp as a First-Class Citizen: TAPT elevates timestamps from auxiliary information to a core component of the recommendation process, enabling a more holistic view of user behavior.

Key Results

  • Comparable or Superior POI Recommendation: TAPT achieved performance comparable to or better than state-of-the-art sequential POI recommendation baselines across the evaluated datasets.
  • Accurate Timestamp Prediction: The model demonstrated the ability to predict the timestamp of the next check-in, with MAE values presented in Table 3. While not directly compared to other timestamp prediction methods (as few exist), the results show promise.
  • Robustness to Parameter Choices: Sensitivity analysis showed that the model's performance was stable with respect to the timestamp embedding dimension and that the weighting parameter for the dual-task loss could be effectively tuned for optimal performance.
  • Improved POI Representation: Visualizations (Figure 6) indicated that TAPT's learned POI representations were better clustered compared to a baseline, suggesting that the temporal information aids in capturing more meaningful POI similarities.

Limitations and Future Work

  • Timestamp Prediction Accuracy: The authors acknowledge that there is still a "considerable gap in accuracy" for timestamp prediction, attributing this to the inherent randomness in user behavior timestamps within real-world LBSN datasets.
  • Dataset Specificity: While TAPT showed strong performance, the improvement on the Gowalla dataset was particularly notable, suggesting that the effectiveness might depend on the dataset's temporal characteristics.
  • Future Directions: The paper implicitly suggests that further research into more sophisticated timestamp encoding methods and exploring how to better handle the inherent randomness in user behavior could lead to even more accurate timestamp predictions. Investigating the interpretability of the learned temporal patterns could also be a valuable future direction.

Summary updated

Abstract

Sequential point-of-interest (POI) recommendation aims to recommend the next POI for users in accordance with their historical check-in information. However, few attempts treat timestamps of check-ins as a core factor for sequence models, leading to insufficient insight into user behavior and subsequently suboptimal recommendations. To address these limitations, we propose to assign equal importance to both POIs and their timestamps, shifting the point of view to recommend the next POI and predict the corresponding timestamp. Along these lines, we present the Time-Aware POI Recommender with Timestamp Prediction (TAPT), a multi-task learning framework for explainable POI recommendations. Specifically, we begin by decoupling timestamps into multi-dimensional vectors and propose a timestamp encoding module to explicitly encode these vectors. Additionally, we design a specialized timestamp prediction module built on the traditional sequence-based POI recommender backbone, effectively learning the strong correlation between POIs and their corresponding timestamps through these two modules. We evaluated the proposed model with three real-world LBSN datasets and demonstrated that TAPT achieves comparable or superior performance in POI recommendation compared to the baseline backbone. Besides, TAPT can not only recommend the next POI, but predict the corresponding timestamp in the future.
BibTeX
@inproceedings{xu2025where,
  author = {Yuanbo Xu and Hongxu Shen and Yiheng Jiang and En Wang},
  title = {{Where and When: Predict Next POI and Its Explicit Timestamp in Sequential Recommendation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/390},
}
Figure