摘要
位置社交网络(LBSN)用户位置数据的分布不均衡,及某些用户出于对隐私安全的考量刻意隐藏自己部分位置信息等因素加剧了兴趣点(POI)推荐难度。就此本文提出了基于元学习的时空神经常微分方程(ML-ODE)来进行有效的下一个POI推荐。该模型主要是将元学习的思想融入到POI推荐过程中,通过不同任务训练优化初始参数,将数据丰富城市中的泛化移动模式迁移到数据匮乏城市,达到优化POI预测任务的目的。该模型将神经常微分方程用于POI推荐领域,定义连续的动态过程,可以接受任意时刻的输入数据,克服了大多数时序推荐模型静态离散化的时间间隔处理方式,更适用于POI序列推荐任务。在真实公开数据集Foursqure上的实验结果表明,ML-ODE在POI推荐方面比当前主流的POI预测方法在NDCG@N指标上提升了超过10%。
In location-based social networks(LBSNs),the uneven distribution of users’ check-in data and the fact that some users deliberately hide some of their location information for privacy and security concerns aggravate the difficulty of point of interest(POI) recommendation.Aiming at this issue,a meta-learning based ordinary differential neural network(ML-ODE) is proposed to carry out effective POI recommendation tasks.ML-ODE leverages the meta-learning mechanism to optimize the parameters of recommendation model.ML-ODE utilizes different tasks to initial parameters,during which the generalized mobility knowledge in the data-rich area is transferred to the datapoor area to achieve the purpose of optimizing the POI recommendation.The network defines a continuous dynamic process that can accept input data that is sporadically-observed.It overcomes the static discretization constraints of most time series prediction models and is more suitable for POI recommendation.The experimental results on the Foursqure real public dataset show that ML-ODE has better performance in POI recommendation than the current state-of-the-art POI recommendation method on the index of NDCG@ N which has been increased by more than10%.
作者
谭海宁
姚迪
毕经平
向徐
杨啸
Tan Haining;Yao Di;Bi Jingping;Xiang Xu;Yang Xiao(University of Chinese Academy of Sciences,Beijing 100049;Institute of Computing Technology,Chinese Academy of Sciences,Beijing 100190;National Key Laboratory of Science and Technology on Blind Signal Processing,Chengdu 610041)
出处
《高技术通讯》
CAS
2021年第12期1248-1260,共13页
Chinese High Technology Letters
基金
国家重点研发计划(2017YFC0820700)
国家自然科学基金(61472403)资助项目。
关键词
兴趣点(POI)推荐
位置社交网络(LBSN)
元学习
神经常微分方程
推荐系统
point of interest(POI)recommendation
location-based social network(LBSN)
meta-learning
neural ordinary differential equation
recommender system
