Sequential Recommendation via Temporal Data Augmentation and Fourier Convolution

Sequence recommendation aims to predict user preferences by modeling users' dynamic historical behaviors. In recent years, advancements in sequence deep learning models such as Transformer have significantly enhanced sequence rec-ommendation. However, most existing sequence recommendation approaches primarily focus on the sequential information of user behaviors while neglecting the critical temporal information. Additionally, noise in interaction records has consistently im-peded the performance of sequence recommendations. To address these two shortcomings, we propose a model named TAFC4Rec. First, to better exploit the relationship between temporal infor-mation and user interest changes, we improve three classical stochastic data augmentation methods by incorporating time interval calculations. Furthermore, following data augmentation, to enhance the encoder's noise filtering capability for better handling augmented data, we design a convolutional structure based on the Fourier transform to denoise the data and capture sequence features. Finally, we conducted experiments on four real-world datasets, and the experimental results demonstrate the effectiveness of our model.