Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Interest Changes: Considering User Interest Life Cycle in Recommendation System

Yinjiang Cai, Jiangpan Hou, Yangping Zhu, Yuan Nie

TL;DR

The study addresses the gap in recommending items by explicitly modeling user interest life-cycles, proposing the Deep Interest Life-cycle Network (DILN) that couples an Interest Life-cycle Encoder Module (ILEM) with an Interest Life-cycle Fusion Module (ILFM). ILEM builds dense representations of emergent, stable, and declining interests by encoding recent activity histograms through a CNN-based Histogram Encoder and a Life-Cycle VQ Cluster. ILFM integrates life-cycle signals into a multi-task ranking framework (MMOE) via a Feature Recalibrator and a Neural Fusion Unit to tailor feature importance and layer-wise representations to the current life-cycle phase. Offline results on public and industrial data show consistent gains in CTR and CVR, and online A/B testing confirms real-world improvements in user engagement and duration, indicating that lifecycle-aware modeling can substantially enhance recommendation systems and is ready for deployment.

Abstract

In recommendation systems, user interests are always in a state of constant flux. Typically, a user interest experiences a emergent phase, a stable phase, and a declining phase, which are referred to as the "user interest life-cycle". Recent papers on user interest modeling have primarily focused on how to compute the correlation between the target item and user's historical behaviors, without thoroughly considering the life-cycle features of user interest. In this paper, we propose an effective method called Deep Interest Life-cycle Network (DILN), which not only captures the interest life-cycle features efficiently, but can also be easily integrated to existing ranking models. DILN contains two key components: Interest Life-cycle Encoder Module constructs historical activity histograms of the user interest and then encodes them into dense representation. Interest Life-cycle Fusion Module injects the encoded dense representation into multiple expert networks, with the aim of enabling the specific phase of interest life-cycle to activate distinct experts. Online A/B testing reveals that DILN achieves significant improvements of +0.38% in CTR, +1.04% in CVR and +0.25% in duration per user, which demonstrates its effectiveness. In addition, DILN inherently increase the exposure of users' emergent and stable interests while decreasing the exposure of declining interests. DILN has been deployed on the Lofter App.

Interest Changes: Considering User Interest Life Cycle in Recommendation System

TL;DR

The study addresses the gap in recommending items by explicitly modeling user interest life-cycles, proposing the Deep Interest Life-cycle Network (DILN) that couples an Interest Life-cycle Encoder Module (ILEM) with an Interest Life-cycle Fusion Module (ILFM). ILEM builds dense representations of emergent, stable, and declining interests by encoding recent activity histograms through a CNN-based Histogram Encoder and a Life-Cycle VQ Cluster. ILFM integrates life-cycle signals into a multi-task ranking framework (MMOE) via a Feature Recalibrator and a Neural Fusion Unit to tailor feature importance and layer-wise representations to the current life-cycle phase. Offline results on public and industrial data show consistent gains in CTR and CVR, and online A/B testing confirms real-world improvements in user engagement and duration, indicating that lifecycle-aware modeling can substantially enhance recommendation systems and is ready for deployment.

Abstract

In recommendation systems, user interests are always in a state of constant flux. Typically, a user interest experiences a emergent phase, a stable phase, and a declining phase, which are referred to as the "user interest life-cycle". Recent papers on user interest modeling have primarily focused on how to compute the correlation between the target item and user's historical behaviors, without thoroughly considering the life-cycle features of user interest. In this paper, we propose an effective method called Deep Interest Life-cycle Network (DILN), which not only captures the interest life-cycle features efficiently, but can also be easily integrated to existing ranking models. DILN contains two key components: Interest Life-cycle Encoder Module constructs historical activity histograms of the user interest and then encodes them into dense representation. Interest Life-cycle Fusion Module injects the encoded dense representation into multiple expert networks, with the aim of enabling the specific phase of interest life-cycle to activate distinct experts. Online A/B testing reveals that DILN achieves significant improvements of +0.38% in CTR, +1.04% in CVR and +0.25% in duration per user, which demonstrates its effectiveness. In addition, DILN inherently increase the exposure of users' emergent and stable interests while decreasing the exposure of declining interests. DILN has been deployed on the Lofter App.
Paper Structure (15 sections, 3 equations, 4 figures, 2 tables)

This paper contains 15 sections, 3 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Interest Life-cycle Encoder Module
  5. Feature engineering
  6. Histogram Encoder
  7. Life-Cycle VQ Cluster
  8. Interest Life-cycle Fusion Module
  9. Feature Recalibrator
  10. Neural Fusion Unit
  11. Experiment
  12. Experimental Settings
  13. Offline Result
  14. Online Result
  15. Conclusion

Figures (4)

  • Figure 1: Comparison between different interest life-cycles. The y-axis represents task-specific efficiency for each interest category.
  • Figure 2: Illustration of feature engineering applied to user exposure history.
  • Figure 3: The architecture of the proposed Deep Interest Life-Cycle Network.
  • Figure 4: Illustration of activated VQ clusters across different interest life-cycles. The y-axis represents the activation probability of each cluster.