DPGN: Denoising Periodic Graph Network for Life Service Recommendation

Different from traditional e-commerce platforms, life service recommender systems provide hundreds of millions of users with daily necessities services such as nearby food ordering. In this scenario, users have instant intentions and living habits, which exhibit a periodic tendency to click or buy products with similar intentions. This can be summarized as the intentional periodicity problem, which was not well-studied in previous works. Existing periodic-related recommenders exploit time-sensitive functions to capture the evolution of user preferences. However, these methods are easily affected by the real noisy signal in life service platforms, wherein the recent noisy signals can mislead the instant intention and living habits modeling. We summarize it as the noise issue. Although there are some denoising recommenders, these methods cannot effectively solve the noise issue for intentional periodicity modeling. To alleviate the issues, we propose a novel Denoising Periodic Graph Network (DPGN) for life service recommendation. First, to alleviate the noisy signals and model the instant intention accurately, we propose (i) temporal pooling (TP) to encode the most representative information shared by recent behaviors; (ii) temporal encoding (TE) to encode the relative time intervals. Second, to capture the user's living habits accurately, we propose the memory mechanism to maintain a series of instant intentions in different time periods. Third, to further capture the intentional periodicity, we propose the temporal graph transformer (TGT) layer to aggregate temporal information. Last, the denoising task is further proposed to alleviate the noisy signals. Extensive experiments on both real-world public and industrial datasets validate the state-of-the-art performance of DPGN. Code is available in https://github.com/ytchx1999/DPGN.