基于GRU神经网络的可移动性兴趣点的推荐系统  

Recommendation System for Mobility Points of Interest Based on GRU Neural Networks

作　　者：史艳翠 张弛[1] SHI Yancui;ZHANG Chi(College of Artificial Intelligence,Tianjin University of Science&Technology,Tianjin 300457,China)

机构地区：[1]天津科技大学人工智能学院,天津300457

出　　处：《天津科技大学学报》2022年第6期54-62,共9页Journal of Tianjin University of Science & Technology

基　　金：国家自然科学基金青年基金项目(61402331,61702367,61807024);天津市教委理工类基本科研业务费项目(2018KJ105)。

摘　　要：目前,兴趣点推荐算法的研究主要聚焦在静态兴趣点的推荐上,对可移动性兴趣点的推荐研究比较少.本文提出了一种包括3个网络模块的复合神经网络结构,可以实现可移动性兴趣点的推荐.整体算法框架包含门控循环(gaterecurrentunit,GRU)神经网络专家模块和门网决策模块,门控循环神经网络专家模块由模式GRU和点GRU两个子模块组成.模式GRU网络模块融入了迁移学习策略,负责学习其他数据集中的动态时空模式,用于模式预测.点GRU网络模块融入了对比学习策略,以达到扩展目标训练集的样本数量、提高网络泛化能力的目的,用于点预测;门网决策模块负责根据输入样本的形式选择相应的GRU模块的输出作为预测结果,实现神经网络专家决策系统.本文还提出了一种结合Tr AdaBoost+最大期望(expectationmaximization,EM)算法的样本过滤算法,该算法能够从两种学习策略产生的扩容样本中筛选出可信的、可用的样本,用于训练模式GRU以及点GRU网络模块.实验结果表明,与马尔可夫(Markov)模型和高斯混合模型(Gaussian mixture model,GMM)方法相比,本文方法泛化能力强,可用小训练样本集驱动神经网络,产生的预测误差以及预测误差分位数更小.At present,the research on point of interest recommendation algorithm mainly focuses on the recommendation of static points of interest.However,there is little research on the recommendation of mobile points of interest. In this article,a composite neural network structure including three network modules is proposed to realize the recommendation of mobile points of interest. The overall algorithm framework includes the gate recurrent unit(GRU)neural network expert module and the gate network decision module. The gate recurrent unit neural network expert module is composed of two sub modules:mode GRU and point GRU. The mode GRU module integrates the transfer learning strategy and is responsible for learning the dynamic spatiotemporal modes in other data sets for mode prediction;the point GRU module integrates the contrast learning strategy to expand the number of samples in the target training set and improve the generalization ability of the network for point prediction;the gate network decision-making module is responsible for selecting the output of the corresponding GRU module as the prediction result according to the form of input samples to realize the neural network expert decisionmaking system. In this article we also propose a sample filtering algorithm combined with TrAdaBoost+expectation maximization(EM). The algorithm can select credible and available samples from the expanded samples generated by the two learning strategies for training mode GRU and point GRU modules. The experimental results show that,compared with Markov model and Gaussian mixture model(GMM)methods,this method has strong generalization ability,can drive the neural network with small training sample set,and produces smaller prediction error and prediction error quantile.

关 键 词：移动性兴趣点 动态时空推荐 迁移学习 对比学习 门控循环神经元 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]