MCPD:结合预训练与去噪图卷积网络的多任务学习推荐模型  

MCPD:Multi-task Learning Recommender System Combining Pre training and Denoising Graph Convolutional Network

作　　者：范钰敏 袁卫华[1] 王龙霄 孙倩 FAN Yumin;YUAN Weihua;WANG Longxiao;SUN Qian(School of Computer Science and Technology,Shandong Jianzhu University,Ji'nan 250100,China;College of Science,City University of Hong Kong,Hong Kong 999077,China)

机构地区：[1]山东建筑大学计算机科学与技术学院,山东济南250100 [2]香港城市大学理学院,中国香港999077

出　　处：《软件导刊》2025年第3期78-85,共8页Software Guide

基　　金：国家自然科学基金项目(62176142,62177031);山东省自然科学基金项目(ZR2021MF099,ZR2022MF334);山东省本科教学改革研究项目(M2021130,M2022245);山东省研究生优质教育教学资源项目(SDYAL2022155);济南市市校融合发展战略工程项目(JNSX2023064)。

摘　　要：目前,基于图卷积网络(GCN)的推荐系统普遍存在噪声、训练效率低以及无法选择合适的损失函数进行有效联合优化的问题。为此,提出结合预训练与去噪图卷积网络的多任务学习推荐模型MCPD,采用图卷积网络重点关注高阶邻居间的协作信号以生成更精准的用户和项目嵌入。首先,通过双向注意力聚合模型分别对用户与项目进行预训练,以提升模型收敛速度和训练效率。其次,设计邻居边去噪自编码器模型,在邻居边去噪任务中将传统图卷积网络与注意力机制相结合以识别噪声边,通过降噪自编码器DAE对嵌入进行编码和解码以减少噪声。最后,选择性能最好的余弦对比损失函数,并结合多任务学习联合优化双向注意力聚合预训练、邻居边去噪和降噪自编码器来保证模型推荐精度。在3个标准数据集上的实验表明,MCPD模型的Recall、NDCG指标分别达到7.10、6.00、19.09以及5.85、4.82、15.75,优于其他基线,在推荐准确性方面相较基于GCN的推荐系统具有明显优势。Currently,recommendation systems based on Graph Convolutional Networks(GCNs)commonly suffer from issues such as noise,low training efficiency,and inability to select appropriate loss functions for effective joint optimization.To this end,a multi task learning recommendation model MCPD is proposed,which combines pre training and denoising graph convolutional networks.The graph convolutional network focuses on the collaborative signals between high-order neighbors to generate more accurate user and item embeddings.Firstly,pre training is conducted on both the user and the project using bidirectional attention to improve the convergence speed and training time efficiency of the model.Secondly,a neighbor edge denoising autoencoder model is designed to combine traditional graph convolutional networks with attention mechanisms in the neighbor edge denoising task to identify noisy edges.The embedding is encoded and decoded using a denoising autoencoder DAE to reduce noise.Finally,select the cosine contrastive loss function with the best performance,and combine multi task learning to jointly optimize bidirectional attention aggregation pre training,neighbor edge denoising,and denoising autoencoder to ensure model recommendation accuracy.Experiments on three standard datasets showed that the Recall and NDCG metrics of the MCPD model reached 7.10,6.00,19.09 and 5.85,4.82,15.75,respectively,outperforming other baselines.In terms of recommendation accuracy,it has significant advantages compared to GCN based recommendation systems.

关 键 词：推荐系统 图卷积网络 协同过滤 去噪 多任务学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]