Abstract: The user-item interaction and other elements in the recommendation algorithm can be mined to form a complete Heterogeneous Information Network (HIN).The traditional recommendation algorithms based on HIN often directly build heterogeneous information networks between users and items,ignoring the similarity between users and items themselves.The constructed networks lack integrity,and the different correlations under different search paths are rarely considered when calculating node correlations.To solve the above problems,a HIN recommendation algorithm considering the similarity between users and items was proposed.Through finding more search paths between the user and the project,and considering different search paths,the Random Walk (RW) in deep learning was introduced to measure the correlation degree between the user and the project node,so as to realize more accurate recommendation.The proposed algorithm was experimented on the Movie Lens dataset,and the experimental results showed that compared with the traditional collaborative filtering recommendation algorithm and the HIN-based recommendation algorithm,the paired random walk algorithm based on different search paths had higher recommendation performance.

Key words: recommendation system, heterogeneous information network, mate path, random walk, HeteSim

摘要： 推荐系统中用户项目之间的交互及其他信息可以构成一个异构信息网络(HIN)。传统基于HIN的推荐算法往往直接构建用户项目间的异构信息网络,忽略了用户用户以及项目项目本身具有的相似性,所构建的网络不够完整,并且在计算节点关联性时鲜有考虑不同搜索路径下的不同关联性。为解决上述问题,提出一种考虑用户及项目本身相似性的HIN推荐算法。通过查找用户与项目之间更多的搜索路径,并考虑不同的搜索路径,引入深度学习中的随机游走(RW)来度量用户项目节点之间的关联度,从而实现更加精确的推荐。将所提算法在公开的Movie Lens数据集上进行了实验,实验结果表明：相较于传统的协同过滤推荐算法以及基于HIN的推荐算法,基于不同搜索路径下成对随机游走的算法具有更高的推荐性能。

关键词: 推荐系统, 异构信息网络, 元路径, 随机游走, HeteSim