Scholarlink

<p>CTR prediction, as a core task in online advertising and e-commerce platforms, directly affects platform revenue and user experience. Although deep learning models have been widely applied in this field, existing research based on the two-stream structure still neglects the potential value of isomorphic networks in feature interaction; mainstream fusion strategies mostly rely on static concatenation or simple weighting, which are unable to dynamically capture high-order semantic correlations across streams. To address these problems, this paper proposes a CTR prediction model called SymmNet based on a symmetric two-stream structure. In terms of feature interaction, a two-stream isomorphic subnetwork structure is adopted, and through differentiated initialization and constraint mechanisms, the feasibility of isomorphic networks in low-cost and high-coverage feature interaction is verified. In terms of fusion strategies, a hierarchical mechanism of “low-level dynamic selection - high-level aggregation alignment”&#xa0;is designed, aiming to break the limitations of static fusion and achieve collaborative optimization of fine-grained features and high-order semantics. Through systematic experiments on two real-world benchmark datasets, the results show that compared with existing mainstream single-stream models, two-stream heterogeneous models, etc., SymmNet achieves statistically significant performance improvements in key evaluation metrics such as AUC and Logloss, verifying its superiority in complex interaction modeling. This study not only provides a new framework of&#xa0;“two-stream isomorphic + hierarchical fusion”&#xa0;for the CTR prediction task, but more importantly, it re-examines the mechanism of the two-stream isomorphic network, providing academic references and empirical evidence for the design of business analytics models.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

SymmNet: A CTR Prediction Model with the Symmetrical Structure

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月2日

查看全文
  • Zilong Jiang,
  • Shachuan Xue,
  • Meiyi Wang,
  • Lin Li,
  • Renlong Jiang,
  • Xiaohui Tao

摘要

CTR prediction, as a core task in online advertising and e-commerce platforms, directly affects platform revenue and user experience. Although deep learning models have been widely applied in this field, existing research based on the two-stream structure still neglects the potential value of isomorphic networks in feature interaction; mainstream fusion strategies mostly rely on static concatenation or simple weighting, which are unable to dynamically capture high-order semantic correlations across streams. To address these problems, this paper proposes a CTR prediction model called SymmNet based on a symmetric two-stream structure. In terms of feature interaction, a two-stream isomorphic subnetwork structure is adopted, and through differentiated initialization and constraint mechanisms, the feasibility of isomorphic networks in low-cost and high-coverage feature interaction is verified. In terms of fusion strategies, a hierarchical mechanism of “low-level dynamic selection - high-level aggregation alignment” is designed, aiming to break the limitations of static fusion and achieve collaborative optimization of fine-grained features and high-order semantics. Through systematic experiments on two real-world benchmark datasets, the results show that compared with existing mainstream single-stream models, two-stream heterogeneous models, etc., SymmNet achieves statistically significant performance improvements in key evaluation metrics such as AUC and Logloss, verifying its superiority in complex interaction modeling. This study not only provides a new framework of “two-stream isomorphic + hierarchical fusion” for the CTR prediction task, but more importantly, it re-examines the mechanism of the two-stream isomorphic network, providing academic references and empirical evidence for the design of business analytics models.