Enhancing Time Awareness in Generative Recommendation

Sunkyung Lee, Seongmin Park, Jonghyo Kim, Mincheol Yoon, Jongwuk LeeEMNLP 2025

TL;DR

This paper introduces GRUT, a novel generative recommendation model that explicitly incorporates temporal dynamics. By leveraging time-aware prompting and trend-aware inference, GRUT significantly improves recommendation accuracy, especially for long-term user preferences, outperforming state-of-the-art models.

What Problem Does This Paper Address?

Traditional generative recommendation (GR) models, while powerful, often overlook the crucial aspect of temporal dynamics in user preferences. They tend to focus on the sequential order of items but fail to capture how user interests evolve over time. This can lead to outdated recommendations, such as suggesting items similar to past purchases even when the user's interests have shifted. For instance, a user who previously bought stuffed animals might now be interested in LEGO products, but a time-agnostic model might still recommend more stuffed animals. This paper aims to bridge this gap by developing a GR model that can effectively understand and leverage temporal signals to provide more timely and relevant recommendations.

Methods

The core innovation of GRUT lies in its Time-aware Prompting mechanism, which injects temporal information into the generative process. This prompting consists of two key contexts:

  • User-level Temporal Context ($C_u$): This captures individual user temporal patterns by encoding absolute timestamps and relative time intervals between interactions. This helps highlight recent interests or stable long-term preferences.
  • Item-level Transition Context ($C_v$): This models collective consumption behaviors by analyzing item transition patterns across all users. It leverages a time-weighted transition graph to identify common sequential item purchases and translates these into natural language prompts.

Following this, Context-integrated ID Generation aggregates these temporal contexts to generate target item IDs.

Additionally, GRUT introduces Trend-aware Inference, a training-free method that refines recommendation rankings by incorporating real-time item trends. This allows the model to adapt to dynamic patterns that emerge after training, enhancing timeliness without requiring retraining. The final score is a combination of the generative model's beam score and the trend score.

Datasets

The authors evaluated GRUT on four real-world datasets:

  • Amazon Review Dataset (three subcategories): "Sports and Outdoors", "Beauty", and "Toys and Games".
  • Yelp Dataset.

These datasets were preprocessed using standard 5-core filtering.

Key Insights and Ideas

  • Temporal Dynamics are Crucial for GR: The paper strongly emphasizes that temporal information is not just an add-on but a fundamental component for effective generative recommendation, significantly outperforming models that ignore it.
  • Complementary Nature of Temporal Signals: Combining absolute timestamps and relative time intervals (user-level context) provides richer information than either signal alone, leading to better performance.
  • Bridging Individual and Collective Temporal Patterns: GRUT successfully integrates both user-specific temporal evolution and broader item transition trends, demonstrating the power of a holistic temporal approach.
  • Trend-aware Inference for Timeliness: The ability to incorporate real-time trends without retraining is a significant practical advantage, allowing recommendations to stay relevant in dynamic environments.
  • Robustness to ID Schemes: GRUT's performance is shown to be robust across different item ID representations (e.g., numeric IDs, title-based IDs), highlighting the strength of its temporal integration.

Key Results

  • State-of-the-Art Performance: GRUT achieves state-of-the-art results across four benchmark datasets, with significant gains of up to 15.4% in Recall@5 and 14.3% in NDCG@5 compared to existing models.
  • Outperforming Baselines: GRUT outperforms the best temporal baseline by 30.8% in R@5 and the best generative baseline by 15.4% in R@5.
  • Long-Term Preference Capture: GRUT shows substantial gains (32.6–46.0% in R@5) in "Long interval groups" (scenarios with long time gaps between interactions), demonstrating its effectiveness in capturing evolving user preferences.
  • Trend-aware Inference Benefits: Applying trend-aware inference to existing generative recommenders leads to average gains of 4.0% in R@5 and 2.9% in N@5, showcasing its generalizability.
  • Minimal Computational Overhead: Trend-aware inference adds only a marginal online latency (1.6-2.2 ms per user), making it practically feasible.

Limitations and Future Work

  • Global Item Transition Graph Noise: The current approach of including all transition pairs in the global item transition graph might incorporate noise or spurious patterns (e.g., accidental clicks). Future work could explore denoising techniques to extract more meaningful temporal patterns.
  • Uniform Temporal Information: The method currently applies temporal information uniformly across all users. However, users exhibit diverse purchasing patterns. Future work could explore user-adaptive modeling, potentially adjusting parameters like $\lambda$ dynamically for individual users.
  • User-Adaptive Trend Influence: The paper suggests that modeling user preferences in a user-adaptive manner would be beneficial, particularly for trend-aware inference, where $\lambda$ could be dynamically adjusted.

Summary updated

BibTeX
@inproceedings{lee2025enhancing,
  author = {Sunkyung Lee and Seongmin Park and Jonghyo Kim and Mincheol Yoon and Jongwuk Lee},
  title = {{Enhancing Time Awareness in Generative Recommendation}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure