基于一维卷积神经网络的序列推荐算法  

Sequence Recommendation Algorithm Based on One-dimensional Convolutional Neural Network

作　　者：黄康鹏 冯锋 HUANG Kang-peng;FENG Feng(School of Information Engineering,Ningxia University,Yinchuan 750021,China)

机构地区：[1]宁夏大学信息工程学院,宁夏银川750021

出　　处：《计算机技术与发展》2025年第3期172-178,共7页Computer Technology and Development

基　　金：宁夏重点研发计划重点项目(2022BEG02016);宁夏自然科学基金项目(2023AAC03031)。

摘　　要：近年来,针对信息过载问题,推荐算法已成为关键解决方法之一,能够有效地提供用户个性化内容。在序列推荐研究中,卷积神经网络因其能有效提取序列信息中的局部特征而受到广泛关注。然而,卷积神经网络在捕捉时序信息方面存在局限性。为了解决这一问题,提出了一种基于一维卷积的序列推荐算法。该算法首先通过卷积操作提取序列的局部特征,然后通过池化操作提取序列的长期特征,并将两者进行加权相加获得用户特征信息,使得其能够综合提取局部特征和长期特征。随后,将用户特征信息与线性变化的序列信息进行点乘,以引入时序信息。此外,将用户特征信息通过前馈网络,以实现非线性转换和跨维度交互增强。最后,对用户特征向量与项目特征向量进行计算获得相关性,并以此进行推荐。实验结果表明,在MovieLens电影数据集和KuaiRand短视频数据集上的测试中,该算法在推荐命中率和归一化折损累计增益等指标上相比四种基线算法均有显著提升,表明该算法能够更有效地进行推荐。In recent years,recommendation algorithms have become a key solution to the problem of information overload,providing users with personalized content effectively.In the field of sequential recommendation,convolutional neural networks have gained widespread attention for their ability to extract local features from sequential information.However,convolutional neural networks have limitations in capturing temporal information.To address this issue,we propose a temporal recommendation algorithm based on one-dimensional convolution.The proposed algorithm first extracts local features from the sequence through convolution operations,then extracts long-term features through pooling operations,and combines the two using weighted addition to effectively capture both local and long-term features.Next,the extracted information is multiplied pointwise with linearly transformed sequential information to introduce temporal information.Additionally,a feedforward network is used to achieve nonlinear transformation and enhance cross-dimensional interactions.Finally,the algorithm calculates the correlation between user feature vectors and item feature vectors to make recommendations.Experimental results show that in tests on the MovieLens movie dataset and the KuaiRand short video dataset,the proposed algorithm significantly improves metrics such as hit rate and normalized discounted cumulative gain compared to four baseline algorithms.It is demonstrated that the proposed algorithm is more effective in making recommendations.

关 键 词：推荐算法 序列推荐 卷积神经网络 前馈网络 用户特征 

分 类 号：TP391.3[自动化与计算机技术—计算机应用技术]