Research On User Behavior Sequence Modeling For Recommender System

Users' behavior data form sequences according to the timestamp of each interaction,which reflect users' changing preferences over time.The sequential recommender systems(SR)learn users' evolving interests based on the behavior sequence and mainly focus on(1)when to recommend and(2)what to recommend next.First,for what to recommend,graph neural networks(GNNs)have gained impressive success in the task of sequential recommendation due to their advantage in obtaining the complex transition patterns of items.However,existing GNN-based sequential recommenders still face some problems.Firstly,the global order is lost when transforming a sequence into a graph.Secondly,the long-term dependencies in a sequence are ignored due to the over-smoothing problem in GNNs.Second,for when to recommend,most previous researches focus on what to recommend but ignore when to recommend.Recommending the right item at the wrong time may waste marketing resources and spread poorly timed advertisements around customers.The right-in-time recommendation task aims to recommend the right item at a proper time.As users'interaction with a specific item is analogous to an event,it is a natural fit to apply the TPP models in the right-in-time sequential recommendation task for the next-item and next-time prediction task.However,with a huge number of items in the recommendation scenario,the temporal point process model faces the unrealistic computation complexities and the inaccurate time prediction issues.For the above problems,in this paper,we give a feasible solution to the problems in the sequential recommender system.The main research contributions are:·For what to recommend,we propose an Order-Aware Graph neural network with Long-range Connections(OAG-LC)for sequence modeling.To efficiently capture the long-term dependencies in a sequence,we construct a sequence into a directed graph via reachability and treat the repeated items in a sequence differently from their positions.To handle the lossy global order problem of GNN,we recurrently encode a sequence into a graph and use a gating mechanism to obtain both order and structure information.Compared to the state-of-the-art models,OAG-LC achieves the best performance on four public dataset and improves more than%3 compared with the baselines.·For when to recommend,we propose a TPP Framework with Self-Adversarial Noise Contrastive Estimation for Right-in-Time Sequential Recommendation,and we call it SaNCE-TSR for short.Our model eliminates the gap between TPP and SR.To solve the unrealistic computation complexities problem,we propose a selfadversarial noise contrastive estimation method for model training.And we design the adversarial sampling mechanism to adaptively sample hard negative items according to the current model.To predict the next time accurately,we predict time in an auto-regressive way by equipping two time-aware sequence encoders with an iterative time prediction module.We conduct experiments on five public dataset,compared to the state-of-the-art baselines,our model achieves the best performance on both of the next item and time prediction task and improves more than%3 compared with the baselines.