A Novel Sequential Recommendation Model Based on the Filter and Model Augmentation

Tianci Yu, Jianxia Chen

Abstract: The sequential recommendation aims to capture the user's dynamic interest characteristics according to the historical interaction sequences of users, to provide users with much more personalized recommendations. Although the popular transformer-based sequential recommendation models can effectively capture items dependencies between sequences, they also have some problems such as the noisy information and data sparseness in reality. Therefore, this paper proposes a novel sequential recommendation model based on Filter and Model Augmentation Strategies, named FMAS in short. In particular, during the embedding process, the proposed FMAS utilizes the Fast Fourier Transform (FFT) to convert the original user sequence embedding into the style of the frequency domain, which can be filtered the noise information via a learnable filter without destroying the sequence correlations; Moreover, to alleviate the problem of data sparsity, the FMAS proposes two model augmentation strategies, named Layer Drop and neural masking, to construct contrastive learning views. Experimental results demonstrate that the FMAS model outperforms many sequential recommendation approaches.