基于多模态相似度增强的点击率预估模型

Click-Through Rate Prediction Model Based on Multimodal Similarity

点击率预估是推荐系统中一个至关重要的任务,直接决定了推荐系统的性能。在点击率预估模型中加入用户的行为序列可以极大地提高点击率预估模型的精度。然而,现有的点击率预估模型通常未考虑用户在文本和视觉方面的信息偏好,这将导致模型难以获得全面的用户兴趣表示,进而使得模型无法获得最优的精度。文章提出了一个统一的多模态用户行为序列建模模型(MSIB:Multi-model similarity improving behavior modeling),使用统一的跨模态预训练框架进行特征抽取,并采用一个多模态相似度增强的注意力机制刻画用户的多模态兴趣偏好。本文方法的有效性通过在真实场景中的大规模数据集下的实验得到了验证。

Click-through rate prediction is a crucial task in recommender systems,which directly determines the performance of recommender systems.Incorporating the user s behavioral se-qucnces into the click-through ratc prcdiction model can grcatly improvc the accuracy.Howev-cr,the cxisting click-through rate prediction models usually do not take into account the user's textual and visual information preferences,which makes it difficult to obtain a comprehensive representation of the user's interests,and thus prevents the model from obtaining optimal accu-racy.In this paper,we propose a unified multimodal user behavior sequence modeling model(MSIB:Multi-model similarity improving behavior modeling)that uses a unified cross-modal prc-training framework for fcaturc cxtraction and a multimodal similarity-cnhanccd attentional mechanism for portraying users multimodal interest preferences.The effectiveness of the meth-od in this paper is validated by experiments under large-scale datasets in real scenarios.