Time-Sensitive Recommendation From Recurrent User Activities

Nan Du, Yichen Wang, Niao He, Jimeng Sun, Le SongNeurIPS 2015 · OpenReview

Introduction

By making personalized suggestions, recommender systems play a crucial role in improving user engagement across modern web services. However, most recommendation algorithms are "static"—they predict a user's preference (like a 1-5 star rating) but ignore the dimension of time. This paper introduces a novel framework for time-sensitive recommendations by combining self-exciting Hawkes processes with low-rank matrix factorization. By modeling user-item interactions as temporal point processes, the system can predict not just what a user wants, but exactly when they are likely to return, scaling efficiently to millions of events.

The Challenge of Time in Recommendations

Real-world behavior is inherently recurrent: we buy groceries every week, listen to favorite songs repeatedly, or experience recurring health issues. The authors identify two major gaps in existing research:

  1. The "When" Problem: How do we predict the next time a user will return to a service?
  2. The "Right Moment" Problem: How do we recommend the most desirable item at a specific moment in time?
Recurrent events and the user-item-time model

Traditional methods like Matrix Factorization or Tensor Factorization often fail here because they don't capture the "self-exciting" nature of human behavior—where performing an action now increases the likelihood of performing it again in the near future. Furthermore, traditional survival analysis (often used for "time-to-event" prediction) assumes events are independent, missing the dependency between past and future actions.

Methods: Low-Rank Hawkes Processes

To address these gaps, the authors propose Low-Rank Hawkes Processes. Here is the breakdown of the technical approach:

  • Hawkes Processes: Unlike a standard Poisson process (which assumes events happen at a constant rate), a Hawkes process is "self-exciting." Every time a user interacts with an item, it creates a "spike" in the intensity function, making another interaction more likely soon after, before decaying over time. This effectively splits user preference into a baseline intensity (long-term preference) and a triggering kernel (short-term, history-dependent preference).
  • Low-Rank Modeling: To handle the "cold start" problem and transfer knowledge between users, the authors assume that the parameters governing these processes live in low-rank matrices. This implies that users and items can be categorized into a few prototypical "types," allowing the model to predict return times even for user-item pairs with very few prior interactions.
  • Hybrid Optimization: Solving for low-rank structures usually requires Singular Value Decomposition (SVD), which is computationally expensive. The authors developed a new algorithm blending Proximal Gradient and Conditional Gradient/Frank-Wolfe methods. Formulating the problem as a convex optimization ensures they can find the global optimum efficiently, avoiding the local minima issues that plague many deep learning-based recommenders.

Datasets

The model was evaluated on both synthetic data and three diverse real-world datasets:

  • last.fm: Music streaming logs (1,000 users, 3,000 artists, ~1 million events).
  • tmall.com: E-commerce shopping events (~26k users, ~2.5k stores).
  • MIMIC II: A medical dataset containing clinical visit records (650 patients, 204 diseases) used to predict the recurrence of specific medical conditions.

Key Results

The proposed model achieved superb predictive performance across multiple time-sensitive recommendation tasks.

Performance results showing convergence, scalability, and prediction errors
  • Predictive Accuracy: The model significantly outperformed baselines like SVD, Poisson processes, and Tensor Factorization. In the last.fm dataset, the Mean Absolute Error (MAE) for item ranking was roughly half that of the next best competitor.
  • Return-Time Prediction: For predicting when a user would return, the Hawkes model achieved the lowest error across all datasets, particularly excelling in the "Heldout" group (users/items not seen frequently in training).
  • Scalability: The algorithm scales linearly with the number of entries, making it feasible for industrial-scale applications involving hundreds of millions of temporal events.
  • Distribution Fit: Using quantile plots, the authors showed that their model's predicted event distribution closely matched actual real-world data, whereas Poisson and Rayleigh distributions deviated significantly.

Limitations and Future Work

While highly effective, the framework has a few limitations. The paper primarily uses an exponential decay for the triggering kernel (the "memory" of the process), but other parametric forms (like Weibull) could be explored for different domains. Additionally, predicting the exact return time requires "Ogata’s thinning algorithm" (a sampling method) because the expectation lacks a simple analytic form, adding a layer of stochastic approximation. Finally, while the framework allows for the inclusion of contextual features like user profiles or location, future work is needed to fully explore these integrations.

Summary updated

Abstract

By making personalized suggestions, a recommender system is playing a crucial role in improving the engagement of users in modern web-services. However, most recommendation algorithms do not explicitly take into account the temporal behavior and the recurrent activities of users. Two central but less explored questions are how to recommend the most desirable item \emph{at the right moment}, and how to predict \emph{the next returning time} of a user to a service. To address these questions, we propose a novel framework which connects self-exciting point processes and low-rank models to capture the recurrent temporal patterns in a large collection of user-item consumption pairs. We show that the parameters of the model can be estimated via a convex optimization, and furthermore, we develop an efficient algorithm that maintains $O(1 / \epsilon)$ convergence rate, scales up to problems with millions of user-item pairs and thousands of millions of temporal events. Compared to other state-of-the-arts in both synthetic and real datasets, our model achieves superb predictive performance in the two time-sensitive recommendation questions. Finally, we point out that our formulation can incorporate other extra context information of users, such as profile, textual and spatial features.
BibTeX
@inproceedings{du2015time,
  author = {Nan Du and Yichen Wang and Niao He and Jimeng Sun and Le Song},
  title = {{Time-Sensitive Recommendation From Recurrent User Activities}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2015},
  url = {https://openreview.net/forum?id=H1EB0v-dWr},
}
Figure