基于对比学习和傅里叶变换的序列推荐算法

Sequence Recommendation Based on Contrast Learning and Fourier Transform

提出一种基于自注意力机制和傅立叶变换的序列推荐算法CSFTRec。通过过滤原始数据中的噪声,最大限度地提高自注意力机制对序列数据的特征捕捉能力。根据对比学习的特点,在贝叶斯个性化排名的基础上引入一种新的对比损失,用于联合训练,可以缩短不同相似序列之间的距离。在8个公共数据集上的实验表明,CSFTRec的收敛速度更快,推荐精度有3%~5%的提高,更适合处理序列数据。

This paper proposes a sequence recommendation algorithm based on self-attention mechanism and Fourier transform,named CSFTRec.By filtering the noise in the original data,this algorithm maximizes the feature capturing ability of the self-attention mechanism on the sequence data.According to the characteristics of contrast learning,a new contrast loss is introduced on the basis of Bayesian personalized ranking for joint training,which can shorten the distance between different similar sequences.Experiments on eight public data sets show that CSFTRec converges faster and improves the recommendation accuracy by 3%to 5%,which indicates that CSFTRec is more suitable for processing sequence data.