基于异质网络嵌入的学术论文推荐方法

Academic paper recommendation based on heterogeneous network embedding

针对基于异质信息网络推荐中的有效信息提取与利用,提出了一种基于异质网络嵌入的学术论文推荐方法。使用由元路径引导的随机游走策略生成节点序列;对于每个元路径,通过最大化序列中相邻节点的共现概率来学习节点的唯一嵌入表示;设计了不同的融合函数,将节点在多个不同元路径的低维表示融合为异质信息网络的嵌入,并且引入注意力机制应用于推荐系统。该方法解决了大多数基于异质信息网络的推荐方法因依赖于基于路径的相似性而无法完全挖掘用户和项目潜在结构特征的问题,在DBLP数据集中验证了模型的有效性,并在RMSE指标中取得超过传统模型的效果。

This paper focuses on the extraction and utilization of information from heterogeneous information network recommendation,and proposes a heterogeneous network embedding-based academic paper recommendation method.The proposed method uses the random walk strategy guided by the meta-path to generate the sequence of nodes.Then,for each meta-path,the unique embedding representation of nodes is learned by maximizing its co-occurrence probability with the adjacent nodes in the sequence obtained from the given meta-path.In addition,in order to fuse the low-dimensional representation of nodes in differentmetapaths into the final output of heterogeneous information network embedding,the different fusion functions are designed to achieve this goal,and attention mechanism is introduced to recommendation system.This paper applies the heterogeneous information network based on meta paths to solve the problem that most recommendation methods based on heterogeneous information networks rely on path similarity,which can not fully mine the potential structural characteristics of users and projects.We use DBLP data set to validate the effectiveness of the proposed model,which achieved better performance compared with the traditional model on RMSE.