HFNF: learning a hybrid Fourier neural filter with a heterogeneous loss for sequential recommendation

Sequential recommendation predicts users’ future interactions by capturing dynamic sequential patterns hidden in their historical behavioral sequences. Recently, many deep neural networks have been applied to learn representations of these sequences. However, noisy interactions in user behavior data impose a negative effect on these learned representations. To address this issue, we propose a model called Hybrid Fourier Neural Filter (HFNF) for the sequential recommendation, which is a combination of two kinds of Fourier neural filters and three kinds of objective functions. In HFNF, a global filter network and an adaptive Fourier neural operator are integrated to model dynamic user behaviors by fully exploiting their respective powerful denoising and expressiveness capabilities. Furthermore, a heterogeneous loss is designed to better train HFNF by enhancing preference learning while preserving the uniformity and the alignment of the learned representations. Experiments conducted on eight benchmark datasets demonstrate the superiority of the proposed HFNF model over competing deep neural network based sequential recommendation models.