«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

期数:

2025年10期

页码:

122-130

栏目:

人工智能

出版日期:

2025-10-10

文章信息/Info

Title:

Recommendation Model Based on Behavior Denoising and Multimodal Alignment

文章编号:

1673-629X(2025)10-0122-09

作者:

宗臣1; 袁卫华1; 孟广婷1; 陈宇1; 王洁宁2; 王星1; 时术华3*

1. 山东建筑大学 计算机科学与技术学院,山东济南 250101;

2. 山东建筑大学 建筑城规学院,山东济南 250101;

3. 山东建筑大学 理学院,山东济南 250101

Author(s):

ZONG Chen1; YUAN Wei-hua1; MENG Guang-ting1; CHEN Yu1; WANG Jie-ning2; WANG Xing1; SHI Shu-hua3*

1. School of Computer Science and Technology,Shandong Jianzhu University,Jinan 250101,China;

2. School of Architecture and Urban Planning,Shandong Jianzhu University,Jinan 250101,China;

3. School of Science,Shandong Jianzhu University,Jinan 250101,China

关键词:

推荐系统; 多模态推荐; 行为去噪; 模态对齐; 图卷积神经网络

Keywords:

recommender system; multimodal recommendation; behavioral denoising; modal alignment; graph convolutional neural network

分类号:

TP391.3

DOI:

10.20165/j.cnki.ISSN1673-629X.2025.0106

摘要:

尽管现有的多模态推荐模型已经取得了巨大成功,但仍存在一些限制。首先,物品的模态数据中存在着大量与用户兴趣无关的信息,这些信息在图神经网络节点之间传播时会影响节点表达的区分度并导致推荐性能不佳。其次,模态数据与用户行为数据以各异的方式来刻画用户的偏好,用户不同模态偏好表示之间处于不同的语义空间,由此带来模态间隙(gap)问题,导致模型无法充分挖掘出各模态之间的潜在互补关系。基于此,该文提出了一种基于行为去噪与多模态对齐的推荐模型(BDMA)。首先,BDMA通过流行度感知的图卷积神经网络向节点注入高阶邻居信息。其次,设计了模态去噪编码器,旨在借助用户行为数据中所包含的行为偏好、行为特征来指导用户模态偏好表示的学习,去除物品模态数据中与用户兴趣无关的噪声信息。再次,BDMA通过轻量化自注意力模块深入挖掘用户不同模态偏好之间的内在关系,构造投影函数,将三种模态的用户偏好投影到统一的空间,实现模态之间的对齐、缓解模态间语义间隙。最后,BDMA在三个广泛使用的公开数据集上进行了大量实验,验证了该模型的有效性。

Abstract:

Despite of the great success of existing multimodal recommenders,there are still some limitations. Lots of information irrelevant to the user’s interest exists in the modal data of items,which will propagate among nodes of the graph neural networks. It weakens the differentiation of the node expressions and leads to poor performance. The modal data and the user’s behavioral data portray the user’s preferences in different ways. The representations of user’s different modal preferences are in different semantic spaces,which leads to the problem of modal gaps and prevents the model from fully exploring the potential complementary relationships between different modalities. Based on this,we propose a Recommendation Model Based on Behavior Denoising and Multimodal Alignment(BDMA). Firstly,in BDMA,information from higher-order neighbors is injected into nodes via a popularity-perception graph convolutional neural network. Secondly,a modal denoising encoder is designed to guide the learning of users’ modal preference representation by leveraging the behavioral preferences and characteristics contained in user behavior data,so as to remove the noise information irrelevant to users’ interests. Next, a lightweight self- attention module is designed to mine the intrinsic relationship between user’s different modal preferences,and a projection function is defined to project the three types of user’s preferences onto a unified space and alleviate inter-modal semantic gaps. Finally,the effectiveness of the proposed BDMA is verified by extensive experiments on three public datasets.

常用功能

更新日期/Last Update: 2025-10-10