Modeling Users' Dynamic Preference for Personalized Recommendation

Xin LiuIJCAI 2015

Introduction

Recommendation systems are fundamental to modern digital platforms, yet they struggle with the inherent fluidity of human interest, known as "preference drift." Static models or simple recency-based windows often fail to capture the complex, non-linear ways in which a user's tastes evolve. This paper proposes a novel approach using Gaussian Processes (GP) to model these dynamic shifts at an individual level, moving beyond the limitations of traditional collaborative filtering.

Methodology: The GP-Based Pipeline

The proposed system moves away from standard collaborative filtering's heavy reliance on crowd behavior, focusing instead on building a precise, time-aware profile for each specific user. The process follows a three-stage pipeline designed to handle implicit feedback and evolving interests:

  1. Topic Modeling (LDA): Items are represented as topic distributions using Latent Dirichlet Allocation (LDA) based on associated textual metadata, such as tags or descriptions. This allows the system to capture the "essence" of an item rather than relying on rigid categories.
  2. Preference Evolution via Gaussian Processes: For each user, the history of their interactions with specific topics is treated as a time series. The authors apply GP regression—specifically utilizing a Neural Network Covariance Function—to predict the user's likely topic distribution at a future timestamp ($t_{x+1}$).
  3. Recommendation via Divergence: Candidate items are ranked by calculating the Jensen-Shannon divergence between the predicted future user distribution and the topic distributions of the items. Items with the lowest divergence (highest similarity) are recommended.
The proposed recommendation pipeline using Gaussian Processes to predict future user preference distributions based on historical topic drift.

Experimental Setup and Datasets

The model was evaluated using two real-world datasets containing implicit feedback and timestamps:

  • Delicious: 1,867 users bookmarking over 69,000 URLs, using tags for topic modeling.
  • Last.fm: 1,892 users listening to over 17,000 music artists, again leveraging user-generated tags.

To address data sparsity for new users, the authors incorporated heuristics like Profile Averaging and User-based Collaborative Filtering to supplement the GP-derived preferences.

Key Results and Discussion

The GP-based approach demonstrated substantial improvements over state-of-the-art baselines, including timeSVD++ and WRMF. The use of a non-stationary Neural Network kernel was particularly effective, as it allowed the model to capture abrupt shifts in interest that standard kernels often smooth over.

Performance comparison (Precision and MRR) on Delicious and Last.fm datasets, showing the GP-based approach consistently outperforming traditional baselines.

Quantitatively, the model achieved a Precision improvement of at least 42.46% and a Mean Reciprocal Rank (MRR) improvement of over 66.14% compared to the best-performing baselines. Notably, the system remains effective even with sparse data, requiring as few as 5–10 past interactions to begin providing accurate predictions.

Conclusion and Future Work

By treating user preference as a continuous, evolving distribution modeled through Gaussian Processes, this research provides a robust framework for personalized recommendation that handles the "drift" of human interest. While the $O(n^3)$ complexity of standard GP regression presents a potential bottleneck for extremely active users, the authors suggest variational Bayesian inference as a path toward greater scalability. Future directions include incorporating granular temporal features, such as time-of-day or day-of-the-week patterns, to further refine the predictive accuracy.

Summary updated

BibTeX
@inproceedings{liu2015modeling,
  author = {Xin Liu},
  title = {{Modeling Users' Dynamic Preference for Personalized Recommendation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2015},
}
Figure