基于随机游走的网络表示学习推荐算法  

Recommendation Algorithm Using Network Representation Learning Based on Random Walk

作　　者：刘峰 王宝亮 邹荣宇 赵浩淳 LIU Feng;WANG Baoliang;ZOU Rongyu;ZHAO Haochun(Information and Network Center,Tianjin University,Tianjin 300072,China;School of Electrical and Information Engineering,Tianjin University,Tianjin 300072,China;International Engineering Institute,Tianjin University,Tianjin 300072,China)

机构地区：[1]天津大学信息与网络中心,天津300072 [2]天津大学电气自动化与信息工程学院,天津300072 [3]天津大学国际工程师学院,天津300072

出　　处：《计算机工程》2021年第9期90-96,105,共8页Computer Engineering

基　　金：赛尔网络下一代互联网技术创新项目“基于IPv6的分布式计算框架模型研究”(NGII20160206)。

摘　　要：根据网络结构中的连接关系得到节点的向量表示,进而将节点的向量表示应用于推荐算法可有效提升其建模能力。针对推荐系统中的同质网络,提出结合随机游走的网络表示学习推荐算法。以DeepWalk算法为基础,在随机游走过程中根据节点重要性设定节点游走序列数,并设置终止概率以控制游走长度优化采样结果,在网络表示学习过程中将SkipGram模型融合节点属性信息,同时考虑上下文节点离中心节点的距离获得更准确的推荐结果。实验结果表明,该算法相比DeepWalk、Node2vec等算法具有更高的推荐准确度,并且较好地解决了冷启动问题。The connections in a network can be simplified into vectors between nodes,and this vectorized representation can be applied to recommendation algorithm to improve their modeling ability.For the homogeneous networks in recommendation systems,a recommendation algorithm using Network Representation Learning(NRL)based on random walk is proposed.The algorithm is constructed based on improved DeepWalk.In the stage of random walk,the walk sequence number of the nodes is set according to the importance of the nodes.In addition,a probability of ending the walk is set to control the length of walk and optimize the sampling results.In the stage of NRL,the node attribute information is fused with the SkipGram model,and the distance between the context node and the central node is considered to improve the accuracy of recommendation results.Experimental results show that the proposed algorithm displays higher recommendation accuracy than DeepWalk,Node2vec and other algorithms.It also provides a solution to the cold-start problem.

关 键 词：推荐算法 网络表示学习 随机游走 序列长度 属性信息 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]