基于自注意力和位置感知图模型的会话推荐  

Session recommendations based on self-attention andposition-aware graph models

作　　者：孙克雷[1] 周志刚[1] SUN Ke-lei;ZHOU Zhi-gang(School of Computer Science and Engineering,Anhui University of Science and Technology,Huainan 232001,China)

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232001

出　　处：《计算机工程与设计》2023年第12期3722-3728,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(61703005)。

摘　　要：为解决现有的会话模型方案都只基于局部会话信息而没有充分考虑全局会话信息的问题,提出一种基于自注意力和位置感知图模型的会话推荐。利用图神经网络构建会话图,利用位置感知注意力建模会话图的一阶邻居信息,引入反向位置嵌入赋予不同项目不同的权重,通过软注意机制获得局部会话表示;利用自注意力机制自适应地捕捉会话的全局依赖;将全局会话与局部会话相结合生成最终会话表示。对3个真实数据集进行实验,模型在3个数据集上P@20分别提升了1.2%、4.3%和12.9%,MRR@20分别提升了2.3%、5.4%和14.3%,验证了所提模型的有效性。To solve the problem that the existing session model scenarios are based only on the local session information and do not adequately consider global session information,session recommendation based on self-attention and position-aware graph model was proposed.The graph neural network was used to construct the session graph,the position-aware attention was used to model the first-order neighbor information of the session graph,the reverse position embedding was introduced to give diffe-rent weights to different items,and the local session representation was obtained through the soft attention mechanism.Global dependencies of a session were adaptively captured with a self-attention mechanism.The global session was combined with the local session to produce the final session representation.Experiments were conducted on three real-world datasets.P@20 is improved by 1.2%,4.3%and 12.9%using the model on three datasets,respectively,while MRR@20 is increased by 2.3%,5.4%and 14.3%respectively,demonstrating the effectiveness of the proposed model.

关 键 词：会话推荐 图神经网络 自注意力机制 反向位置嵌入 软注意力机制 邻居信息 位置感知图模型 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]