|
| 融合上下文感知的深度残差点击率预测模型 |
| Deep residual click-through rate prediction model incorporating context-awareness |
| 投稿时间：2024-11-18 修订日期：2024-12-23 |
| DOI： |
| 中文关键词: 关键词: 点击率预测 推荐系统 上下文感知 特征交互 深度学习 |
| 英文关键词: Keywords: click-through prediction recommender systems context-awareness feature interaction deep learning |
| 基金项目:国家自然科学基金(62001004)，安徽省自然科学基金(2008085MF218)，安徽省高校优秀青年人才支持计划重点项目(gxyqZD2021124)，安徽省高校省级自然科学研究重大项目(2024AH040039) |
|
| 摘要点击次数: 154 |
| 全文下载次数: 0 |
| 中文摘要: |
| 摘 要: 点击率（Click-Through Rate，CTR）预测在广告和电子商务领域应用广泛，众多点击率预测模型应运而生。然而，已有的CTR预测模型大多只关注单一特征的固定表示，忽视了每个特征在不同上下文中的不同重要性，导致模型的性能不佳。此外，现有模型在高阶特征交互与细粒度特征融合方面存在不足，难以提升模型的表达能力。为解决上述问题，提出了一种融合上下文感知的深度残差点击率预测模型(Context-aware Deep Residual, CDR)。首先，该模型通过上下文聚合单元CAU捕获上下文相关信息以及特征之间的关系信息，生成上下文感知特征以丰富特征表示；其次，通过将残差连接与MLP网络相结合，以优化特征交互的非线性变换，增强模型对高阶特征的学习能力；最后，利用双线性融合操作实现更加细粒度的特征融合，提升了特征表示的全面性与鲁棒性。在Criteo、Avazu、Movielens和Frappe等公开数据集上进行了对比实验，AUC指标平均提升了1.04%，Logloss指标平均改善了2.27%。结果表明，该模型的性能优于现有先进模型，有效提升了CTR预测的精度。 |
| 英文摘要: |
| Abstract: Click-Through Rate prediction is widely used in advertising and e-commerce, and numerous click-through rate prediction models have emerged. However, most of the existing CTR prediction models only focus on the fixed representation of a single feature, ignoring the dynamic importance of each feature in different contexts, resulting in poor model performance. In addition, existing models are deficient in high-order feature interaction and fine-grained feature fusion, which makes it difficult to improve the model's expressive ability. To solve the above problems, a context-aware deep residual click-through rate prediction model (Context-aware Deep Residual, CDR) is proposed. First, the model generates context-aware features to enrich the feature representation by capturing context-related information and relationship information between features through the context aggregation unit CAU; second, the model enhances the model's ability to learn higher-order features by combining the residual linkage with the MLP network in order to optimize the nonlinear transformation of feature interactions; and lastly, the model utilizes the bilinear fusion operation to achieve more fine-grained feature fusion. improves the comprehensiveness and robustness of feature representation. Comparison experiments are conducted on public datasets such as Criteo, Avazu, Movielens, and Frappe, and the AUC metrics are improved by an average of 1.04%, and the Logloss metrics are improved by an average of 2.27%. The results show that the model outperforms the existing state-of-the-art models and effectively improves the accuracy of CTR prediction. |
|
View Fulltext
查看/发表评论 下载PDF阅读器 |
| 关闭 |
