基于注意力机制的协同卷积动态推荐网络  

Attention-based Collaborative Convolutional Dynamic Network for Recommendation

作　　者：汤文兵 任正云[1] 韩芳[1] TANG Wen-Bing;REN Zheng-Yun;HAN Fang(College of Information Science and Technology,Donghua University,Shanghai 201620)

机构地区：[1]东华大学信息科学与技术学院,上海201620

出　　处：《自动化学报》2021年第10期2438-2448,共11页Acta Automatica Sinica

基　　金：国家自然科学基金(11572084,11972115)资助。

摘　　要：一直以来,各种推荐系统关注于如何挖掘用户与物品特征间的潜在关联,特征信息的充分利用有利于用户到物品的精准匹配.基于矩阵分解和分解机的推荐算法是该领域的主流,前者学习用户历史行为而后者分析对象特征关系,但都难以兼顾用户行为与个体特征.而近年来,深度神经网络凭借其强大的特征学习能力和灵活可变的结构被应用到了推荐系统领域.鉴于此,本文提出了一种基于注意力机制的协同卷积动态推荐网络(Attentionbased collaborative convolutional dynamic network,ACCDN),它通过注意力机制实现用户历史行为、用户画像与物品属性的多重交互,再通过卷积网络逐层捕捉更高阶的特征交互.网络同时接受不同组块输出的低阶至高阶信息,最后给出用户对指定物品青睐评分概率的预估.而且本文还提出了一种基于无参时间衰减的用户兴趣标签来量化用户关注的变化.通过比较若干先进模型在两个现实数据集的表现,本文设计的动态推荐模型不但能够缓解推荐时滞性,还能明显提高推荐质量,为用户带来更好的个性化服务体验.A variety of recommender systems focus on connecting users′and item′s features,and the full use of the potential information presented by features contributes to the accurate user-item matching.Recommendation algorithms based on matrix factorization and factorization machine have risen to the backbone in this field,with the former learning the users′behavior in history and the latter analyzing the individual features,but neither of them are taken into account by algorithms.Recently,deep neural networks are widely applied to recommendation because of their powerful representation learning and flexible structures.Thus,this paper proposes a novel attentionbased collaborative convolutional dynamic network(ACCDN)for recommendation,which realizes the multiple interactions of the user behavior,user profiles and item attributes through attention mechanism and models higher-order interactions by convolution layers.And it absorbs all the low-to-high-order interaction features to predict the user preferable rating probability given an item.In addition,this paper puts forward the unique user interest label based on non-parametric time decay,as an auxiliary tool,to quantify the change of user focus.We conduct extensive experiments on two real-world datasets,which justifies that our proposed ACCDN not only alleviates the recommendation lag,but also improves the quality of recommendation significantly,bringing a better personalized service experience.

关 键 词：注意力机制 2D 卷积网络 动态推荐 多重特征交互 

分 类 号：TP391.3[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]