Relation-enhance Rec
relation-enhance KG
RE-KGR
Paper: RE-KGR: Relation-Enhanced Knowledge Graph Reasoning for Recommendation
总结：把relation当做向量空间，同时考虑relation的方向性，最后基于路径概率预测
Methodology
given a CKG
Embedding Layer
for every entity and relation, one-hot to dense vector
RGC Layer
First-order Aggregation:
project each entity t to a different semantic space conditioned to the relation r:
Here, \(M_{r−1}\), \(M_r\) are mapping matrices, and r and r−1 are a pair of inverse relations,such as AuthorOf and WrittenBy.
High-order Aggregation:
Here, is the concatenation operator, and e(0) denotes initial embeddings.(dense connectivity)
Local Similarity Layer
Prediction Layer
Predict Based on path:
use \(P_{UIIP}={(h,r,t)|(h,r,t)\in G}\) to describe an acyclic UIIP, the probability of the UIIP is:
we use Pui to denote all acyclic UIIPs that start and end with user u and item i.
PeRN
