Paper title ：

Modeling Scale-free Graphs with Hyperbolic Geometry for Knowledge-aware Recommendation

The conference ：

WSDM 2022

Thesis link ：

https://arxiv.org/abs/2108.06468

Code link ：

https://github.com/yankai-chen/LKGR

In order to alleviate the cold start and data sparsity problems in traditional recommendation systems , In recent years , Introducing external knowledge into recommendation systems to build knowledge maps has attracted more and more attention . Besides , Due to the powerful performance of graph neural network in extracting graph data features , Some studies have linked recommendation systems to GNN Combined . be based on GNN The knowledge map recommendation model usually puts users - The interaction of object history and external knowledge map is unified into three parts , However, after data unification , These tripartite graphs usually exhibit scale-free （ Or hierarchy ） The characteristics of graphs , Pictured 1（a） Shown , The degree distribution of the two benchmark data sets is approximate to the power-law distribution .
The existing research shows that , For trees （ Power law distribution ） data , European space will get higher distortion , similarly , The traditional graph embedding method based on Euclidean space may not be able to effectively capture the internal hierarchy of scale-free networks , Thus, the node embedding is highly distorted , Finally, the recommended performance is reduced .
Existing research shows that hyperbolic space , That is, the continuous tree space with exponential growth characteristics , Data with hierarchical data structure or scale-free network structure can produce less distortion , Pictured 1（b） Shown , In hyperbolic space , The distance between nodes near the center of the graph is small , The distance between nodes near the graph boundary is larger . In order to solve the above problems , This paper proposes a knowledge perception recommendation model based on the hyperbolic geometry Lorentz model , Referred to as LKGR.