Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Time Matters: A Novel Real-Time Long- and Short-term User Interest Model for Click-Through Rate Prediction

Xian-Jin Gui

TL;DR

This work tackles CTR prediction by modeling user interests that evolve with time, introducing two regular patterns: periodic and time-point patterns. It proposes TLSI, a time-aware framework that learns dynamic long- and short-term interests tailored to the target item and prediction time, using content- and time-based attention for long-term signals, a temporally augmented LSTM for short-term signals, and a gated fusion to combine them. The approach is validated on multiple public datasets and a large industrial dataset, showing consistent gains in AUC and logloss over strong baselines and demonstrating the practical value for real-time personalization, particularly in video platforms. The results highlight the importance of explicit time-point and time-interval information in capturing user interest at the moment of prediction, enabling more accurate CTR predictions.

Abstract

Click-Through Rate (CTR) prediction is a core task in online personalization platform. A key step for CTR prediction is to learn accurate user representation to capture their interests. Generally, the interest expressed by a user is time-variant, i.e., a user activates different interests at different time. However, most previous CTR prediction methods overlook the correlation between the activated interest and the occurrence time, resulting in what they actually learn is the mixture of the interests expressed by the user at all time, rather than the real-time interest at the certain prediction time. To capture the correlation between the activated interest and the occurrence time, in this paper we investigate users' interest evolution from the perspective of the whole time line and develop two regular patterns: periodic pattern and time-point pattern. Based on the two patterns, we propose a novel time-aware long- and short-term user interest modeling method to model users' dynamic interests at different time. Extensive experiments on public datasets as well as an industrial dataset verify the effectiveness of exploiting the two patterns and demonstrate the superiority of our proposed method compared with other state-of-the-art ones.

Time Matters: A Novel Real-Time Long- and Short-term User Interest Model for Click-Through Rate Prediction

TL;DR

This work tackles CTR prediction by modeling user interests that evolve with time, introducing two regular patterns: periodic and time-point patterns. It proposes TLSI, a time-aware framework that learns dynamic long- and short-term interests tailored to the target item and prediction time, using content- and time-based attention for long-term signals, a temporally augmented LSTM for short-term signals, and a gated fusion to combine them. The approach is validated on multiple public datasets and a large industrial dataset, showing consistent gains in AUC and logloss over strong baselines and demonstrating the practical value for real-time personalization, particularly in video platforms. The results highlight the importance of explicit time-point and time-interval information in capturing user interest at the moment of prediction, enabling more accurate CTR predictions.

Abstract

Click-Through Rate (CTR) prediction is a core task in online personalization platform. A key step for CTR prediction is to learn accurate user representation to capture their interests. Generally, the interest expressed by a user is time-variant, i.e., a user activates different interests at different time. However, most previous CTR prediction methods overlook the correlation between the activated interest and the occurrence time, resulting in what they actually learn is the mixture of the interests expressed by the user at all time, rather than the real-time interest at the certain prediction time. To capture the correlation between the activated interest and the occurrence time, in this paper we investigate users' interest evolution from the perspective of the whole time line and develop two regular patterns: periodic pattern and time-point pattern. Based on the two patterns, we propose a novel time-aware long- and short-term user interest modeling method to model users' dynamic interests at different time. Extensive experiments on public datasets as well as an industrial dataset verify the effectiveness of exploiting the two patterns and demonstrate the superiority of our proposed method compared with other state-of-the-art ones.

Paper Structure

This paper contains 17 sections, 15 equations, 4 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Temporal Information
  4. Long- and Short-term User Interest
  5. Proposed Method
  6. Temporal Information Extraction
  7. Modeling Long-term Interest
  8. Modeling Short-term Interest
  9. Long- and Short-term Interest Fusion
  10. Training and Inference
  11. Experiments
  12. Evaluation Datasets
  13. Compared Methods
  14. Experiments on Public Datasets
  15. Ablation Study
  16. ...and 2 more sections

Figures (4)

  • Figure 1: An example of the periodic patterns and the time-point patterns.
  • Figure 2: The overall structure of the TLSI model. The core part of the model is the long-term module and the short-term module, which generates long-term interest and short-term interest from a user’s behavioral sequence respectively.
  • Figure 3: The mean values of the gate vector $\alpha_t^u$ for long-term and short-term interest fusion.
  • Figure 4: Distribution of attentive weights (mean and std) of two attention mechanisms for long-term interest module.