Recommendation Algorithms against the Malicious Competition in the Market

In recent years, sequential recommendation tasks have made significant progress with the support of deep learning. Existing recommendation algorithms mainly rely on the historical interaction information between users and items to capture the sequential features of user behaviors and make precise personalized recommendations. However, malicious market competition often influences recommendation decisions through fake user behaviors in real-world scenarios. Previous research has often ignored whether the algorithms can maintain robustness against malicious attacks, which has led to a large gap between research and applications. In this work, we propose the DF-Rec model, which is a self-attentive neural structure with learnable filters for sequential recommendation tasks. The model introduces the concept of dynamic filters, Fourier transform-based filters that effectively achieve attenuation of user behavior data noise in the frequency domain, while allowing our model to have lower time complexity and learnable filters that are robustly adaptive, increasing the model flexibility without over-increasing the number of model parameters. Experiments on the public dataset RecSys Challenge 2017 show that our proposed approach outperforms the widely used CNN and transformer-based methods.