Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Scene-wise Adaptive Network for Dynamic Cold-start Scenes Optimization in CTR Prediction

Wenhao Li, Jie Zhou, Chuan Luo, Chao Tang, Kun Zhang, Shixiong Zhao

TL;DR

SwAN tackles the cold-start problem in dynamic multi-scene CTR prediction by learning scene relationships and user-driven similarity through Scene Relation Graphs and Similarity Attention Networks, and by adaptively composing scene-specific and shared representations via a Differentiable Ensemble-experts Module. The approach combines Cross-scene Feature Representation with a gated Decision Layer to efficiently bootstrap new scenes and reduce negative transfer, validated through industrial Meituan deployment and public Taobao data. It achieves notable gains in CTR and order volume, while ablations confirm the importance of each component (SRG, SAN, CFR, AEM) and the custom loss terms for promoting diversity and robust allocation across scenes. The practical impact includes scalable, online adaptation to hundreds of scenes with manageable latency, making SwAN suitable for rapidly evolving mobile E-commerce environments.

Abstract

In the realm of modern mobile E-commerce, providing users with nearby commercial service recommendations through location-based online services has become increasingly vital. While machine learning approaches have shown promise in multi-scene recommendation, existing methodologies often struggle to address cold-start problems in unprecedented scenes: the increasing diversity of commercial choices, along with the short online lifespan of scenes, give rise to the complexity of effective recommendations in online and dynamic scenes. In this work, we propose Scene-wise Adaptive Network (SwAN), a novel approach that emphasizes high-performance cold-start online recommendations for new scenes. Our approach introduces several crucial capabilities, including scene similarity learning, user-specific scene transition cognition, scene-specific information construction for the new scene, and enhancing the diverged logical information between scenes. We demonstrate SwAN's potential to optimize dynamic multi-scene recommendation problems by effectively online handling cold-start recommendations for any newly arrived scenes. More encouragingly, SwAN has been successfully deployed in Meituan's online catering recommendation service, which serves millions of customers per day, and SwAN has achieved a 5.64% CTR index improvement relative to the baselines and a 5.19% increase in daily order volume proportion.

Scene-wise Adaptive Network for Dynamic Cold-start Scenes Optimization in CTR Prediction

TL;DR

SwAN tackles the cold-start problem in dynamic multi-scene CTR prediction by learning scene relationships and user-driven similarity through Scene Relation Graphs and Similarity Attention Networks, and by adaptively composing scene-specific and shared representations via a Differentiable Ensemble-experts Module. The approach combines Cross-scene Feature Representation with a gated Decision Layer to efficiently bootstrap new scenes and reduce negative transfer, validated through industrial Meituan deployment and public Taobao data. It achieves notable gains in CTR and order volume, while ablations confirm the importance of each component (SRG, SAN, CFR, AEM) and the custom loss terms for promoting diversity and robust allocation across scenes. The practical impact includes scalable, online adaptation to hundreds of scenes with manageable latency, making SwAN suitable for rapidly evolving mobile E-commerce environments.

Abstract

In the realm of modern mobile E-commerce, providing users with nearby commercial service recommendations through location-based online services has become increasingly vital. While machine learning approaches have shown promise in multi-scene recommendation, existing methodologies often struggle to address cold-start problems in unprecedented scenes: the increasing diversity of commercial choices, along with the short online lifespan of scenes, give rise to the complexity of effective recommendations in online and dynamic scenes. In this work, we propose Scene-wise Adaptive Network (SwAN), a novel approach that emphasizes high-performance cold-start online recommendations for new scenes. Our approach introduces several crucial capabilities, including scene similarity learning, user-specific scene transition cognition, scene-specific information construction for the new scene, and enhancing the diverged logical information between scenes. We demonstrate SwAN's potential to optimize dynamic multi-scene recommendation problems by effectively online handling cold-start recommendations for any newly arrived scenes. More encouragingly, SwAN has been successfully deployed in Meituan's online catering recommendation service, which serves millions of customers per day, and SwAN has achieved a 5.64% CTR index improvement relative to the baselines and a 5.19% increase in daily order volume proportion.
Paper Structure (17 sections, 15 equations, 6 figures, 14 tables)

This paper contains 17 sections, 15 equations, 6 figures, 14 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Approach
  4. Scene Relation Graph (SRG)
  5. Similarity Attention Network (SAN)
  6. Cross-scene Feature Representation (CFR)
  7. Adaptive Ensemble-experts Module (AEM)
  8. Decision Layer
  9. Composition of Losses
  10. Experiments
  11. Experimental Settings
  12. Experimental Results
  13. Hyperparameters Experiments
  14. Ablation Study and Analysis
  15. Application in Practice
  16. ...and 2 more sections

Figures (6)

  • Figure 1: The figure displays multiple business scenes' online / offline states over time. The $x$-axis represents time, with green and blue segments indicating the spring and winter. Boxes above the $x$-axis show the online and offline activities during certain periods within each season. Activity diagrams represent online status, while dashed lines represent offline (e.g., the Valentine's Day activity is only online on Tuesdays in the left graph). The findings suggest that scenes go online immediately when an activity starts (cold-start problem) and go offline right after it ends (limited online time and sample accumulation). The actual online business is even more time-sensitive, with over $200$ scenes going online / offline on average each month. This is the dynamic multi-scene problem introduced in this paper, which poses significant challenges to existing multi-scene models.
  • Figure 2: Schematic diagram of the SwAN structure.
  • Figure 3: The relation between different scenes in SRG. The numbers on the lines are the number of the same key features.
  • Figure 4: Similarity Attention Network.
  • Figure 5: The t-SNE dimensionality reduction visualization of item embeddings, where different colors represent different categories in the original data (gray represents missing category information).
  • ...and 1 more figures