作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2023, Vol. 49 ›› Issue (5): 105-111,121. doi: 10.19678/j.issn.1000-3428.0064747

• 人工智能与模式识别 • 上一篇    下一篇

小波卷积增强的对比学习推荐算法

许凤, 杨兴耀, 于炯, 李梓杨, 李晨瑜, 张君   

  1. 新疆大学 软件学院, 乌鲁木齐 830008
  • 收稿日期:2022-05-19 修回日期:2022-07-27 发布日期:2022-09-20
  • 作者简介:许凤(1998-),女,硕士研究生,主研方向为推荐系统;杨兴耀(通信作者),副教授、博士;于炯,教授、博士;李梓杨,副教授、博士;李晨瑜、张君,硕士研究生。
  • 基金资助:
    国家自然科学基金(61862060,61966035,61562086);新疆维吾尔自治区教育厅项目(XJEDU2016S035);新疆大学博士科研启动基金(BS150257);新疆维吾尔自治区自然科学基金(2022D01C56)。

Wavelet Convolution Enhanced Contrastive Learning Recommendation Algorithm

XU Feng, YANG Xingyao, YU Jiong, LI Ziyang, LI Chenyu, ZHANG Jun   

  1. School of Software, Xinjiang University, Urumqi 830008, China
  • Received:2022-05-19 Revised:2022-07-27 Published:2022-09-20

摘要: 推荐算法是一种用于解决信息过载问题的方法,序列化推荐通过建模用户购买的物品序列预测下一个物品。现有的序列化推荐算法通常忽视用户行为序列中的噪声、跨序列信息和物品间的组合依赖等问题,导致推荐性能受限。为此,提出一种小波卷积增强的对比学习推荐算法WCLR。利用数据的内在相关性获得自监督信号,并根据预训练的方法来增强数据表示。给出3个辅助的自监督学习任务,利用信息最大化原理学习属性、物品、序列与邻居序列的相关性,通过互信息最大化提供一种统一的方式描述不同类型数据间的相关性。由于小波卷积网络能提取物品的组合依赖,降低用户交互序列中的噪声,设计一个多核小波卷积模块,通过多尺寸用户序列多方面捕获用户的潜在兴趣,将自监督学习和小波卷积融入到推荐算法模型中,降低序列数据稀疏性和噪声,提高推荐精度。在LastFM、Beauty和Toys 3个数据集上的实验结果表明,与8个序列化推荐模型相比,WCLR算法的命中率、归一化折损累计增益和平均倒数秩分别提升了3.30%、1.47%和2.17%。

关键词: 推荐算法, 序列化推荐, 小波变换, 小波卷积网络, 自监督学习

Abstract: Recommendation algorithms are methods of solving information overload,which use sequential recommendations to predict the next item by modeling the sequence of items purchased by the user.Existing sequential recommendation algorithms ignore the noise and union-level influence between items,which results in limited recommendation performance.Therefore,this study proposes a Wavelet convolution enhanced Contrastive Learning Recommendation(WCLR) algorithm.The model exploits the intrinsic data correlation to obtain self-supervised signals and enhances the data representation using pre-trained methods.Three auxiliary self-supervised learning tasks are designed,which use the principle of information maximization to learn the correlation between the attribute,item, sequence,and neighbor sequence.Mutual Information Maximization(MIM) provides a unified approach to describing the correlation between the different types of data.Wavelet convolution extracts the combinatorial dependencies of the items and reduces the noise in the user sequence.The model captures the potential interest of the users through the multiple aspects of the multi-size user sequences by designing a multi-core wavelet convolution module.The WCLR algorithm integrates self-supervised learning and wavelet convolution into the model to reduce the sparsity and noise of the sequence data and improve recommendation accuracy.The experimental results from the LastFM,Beauty,and Toys data sets show that the Hit Rate(HR),Normalization Discounted Cumulative Gain(NDCG),and Mean Reciprocal Rank(MRR) of the WCLR algorithm model increased by 3.30%,1.47%,and 2.17%,respectively,compared to those of the eight serialized recommended models.

Key words: recommendation algorithm, sequential recommendation, wavelet transform, wavelet convolution network, self-supervised learning

中图分类号: