Research On User Generated Seqence-oriented Deep Learning Technology

For a long time,user-generated sequence modeling has been a research hot spot in recommender systems.Conventional methods usually take lots of manpower and material costs to quickly filter out information that matches user interests from massive data.Thanks to massive data and powerful computing power,deep learning technology has achieved great success in sequential modeling tasks.However,the current research on user-generated sequence modeling still has many defects.For user homogeneous sequence data 1)The modeling of user's temperature sequence and multi-type features in electronic health records is insufficient;2)The effects of next check-in time modeling are ignored.For user heterogeneous sequence data 3)The fusion of two-stage workflow of sequence segmentation and recommendation is not considered in sequential recommendation task.This article conducts in-depth research on the problems that exist in user homogeneous and heterogeneous sequence data,and designs data-driven personalized user sequential modeling frameworks to better alleviate the above problems.The main work and contributions of this paper are summarized as follows:In this paper,We first study the medical data of CAP.We formulate a new problem that automatically detecting pathogenic microorganism of CAP by considering patient biomedical features from EHRs,including time-varying body temperatures and common laboratory measurements.We further develop a Patient Attention based Recurrent Neural Network(PA-RNN)model to fuse different patient features for detection.We conduct comprehensive experiments on a real dataset,indicating the benefits of fusing multiple features in EHRs to the research problem,and demonstrating the effectiveness of PA-RNN over several alternative methods.Most of the existing studies in Location-based social networks(LBSNs)only focus on predicting the spatial aspect of check-ins,yet the joint prediction of the spatial and temporal aspects more meets real application scenarios.To this end,we propose a deep multi-task learning model,ARNPP-GAT,which integrates user long-term representation learning,short-term behavior modeling,and temporal point process into a unified architecture.Specifically,ARNPP-GAT leverages graph attention network to learn the long-term representation of users by encoding their social relations.Short-term preference is revealed in the sequential modeling of users' recent check-ins.Meanwhile,attention-based recurrent neural point process endows the model with the capability of characterizing the effects of past check-in events and performing multi-task learning to predict the locations and timestamps of the next check-ins.Empirical results on two real-world datasets demonstrate ARNPP-GAT is superior compared with several competitors,validating the contributions of multi-task learning and social relation modeling.Finally,the current session-based recommendation is a two-stage workflow consists of session segmentation and session recommendation.To bridge the gap,we devise a novel reinforcement learning framework for session-based recommendation(RL4SRec).This framework enjoys the capability of unifying the two stages in an end-to-end learning fashion,ensuring the session segmentation stage to be guided by the signal from the final recommendation performance.Specifically,a policy network is defined to take an action at each position of a user-interacted sequence in chronological order.The action determines whether to segment the sequence or not in the current position.The policy network is optimized by the accumulated rewards calculated from an adopted sessionbased recommendation model.The recommendation model is in turn trained on the segmented sessions generated by the policy network.It is worth noting that our framework RL4 SRec is flexible and can incorporate existing session-based recommendation models as its component without much effort.We conduct experiments on three commonly used datasets and demonstrate that the learning framework brings considerable improvements to three recommendation models,compared with using carefully selected session segmentation strategies.In summary,this paper designs data-driven user-specific sequential modeling frameworks for multiple types of user-generated data,and performs experiments on multiple real-world datasets.The experimental results prove the effectiveness of the proposed method.