Dual Frequency-based Temporal Sequential Recommendation

Sequential recommendations aim to capture user preferences through user historical behavior interaction data in order to make accurate recommendations. Recently, graph convo-lutional networks have achieved remarkable results in the field of sequential recommendations. However, most of them only utilize the original interactive items, ignoring the influence of time and noise information in their interaction sequences. In particular, some of them may pay attention to the time domain information, they also neglect the frequency domain information which can also be utilized to analyze user interests. To address these limitations, considering both time domain and frequency domain perspectives, we propose a novel model, named DFT-SR in short. First, our approach incorporates a timestamp embedding based on a window function to capture the temporal representations of user interaction sequences. Afterward, we replace the self-attention layer in the encoder with a learnable filter module, which comprises two components such as high-frequency and low-frequency functions, utilizing different neural network layers in the frequency domain to hierarchically cover specific frequency ranges. Experimental results demonstrate the superiority of DFT-SR model over other sequence models. The incorporation of frequency-aware filtering and timestamp embedding enhances the performance of sequential recommendations, increasing the HR@20 from 23.71% to 35.70%, and increasing the NDCG@20 from 26.54% to 51.28%.