Hierarchical attention graph convolution network for interpretable recommendation based on knowledge graph

【 Reading notes 】
Prepare to write a complete note carefully ！！！
Could it be this one ？？？

Periodical KBS The article ：HAGERec: Hierarchical Attention Graph Convolutional Network Incorporating Knowledge Graph for Explainable Recommendation

motivation ：

be based on GCN Most of our recommendations focus on isomorphism diagrams or users - Item similarity diagram , The failure to make full use of heterogeneity is just the complex and rich semantics between entities in the graph . This article ：HAGERec, from isomerism Explore the user's potential preferences in the higher-order connectivity structure of the knowledge graph .

Method ：

Two way information dissemination strategy (item+user)+ Hierarchical attention mechanism
One side ,HAGERec In the process of messaging , By merging the neighborhood information of entities , use GCN Modeling the local adjacent structure of the entity .
On the other hand , structure Hierarchical attention mechanism , Adaptively develop and adjust collaboration signals for each user item pair , User item pairs are divided into neighbor level attention and prediction level attention .

analysis RippleNet：
RippleNet A combination of embedding based and path based methods , Users' hierarchical interests can be automatically captured from high-level connections .
There are certain limitations :(1) Not fully excavated Semantic information of relationships between entities , This leads to the loss of information in the message delivery process .(2) RippleNet A presentation model that focuses on users , None of them Applicable information communication strategies to learn at the same time item Aggregate representation of .(3) If there is no appropriate strategy to sample the relevant neighbor entities , On such a large scale ripple set This will result in higher computation and storage costs .

analysis KGCN：
KGCN Just simply put GCN Apply to... Based KG In my recommendation .
There are some drawbacks :(1) It only focuses on item Modeling of representations , There is no one Applicable information communication strategies to learn at the same time user Aggregate representation of .(2) Nor does it take into account interpretability , Resulting in insufficient understanding of user preferences .

HAGERec The frame diagram of

1. Flatten and embedding layer： Entity and relationship vectorization
   “ flat ” operation , Further make the connection under the relationship orderly , Better distinguish between different connections , Lay the foundation for constructing the connection of knowledge perception .

2. GCN learning layer： polymerization user and item Information about
   Two way entity propagation strategy , Storage and dissemination come from item and user More information about neighbors , So as to better adapt to local adjacent structures .
   

3. Interaction signals unit： Retain the interactive signal structure of the entity and its adjacent network , Make the presentation of users and items more complete .
   

4. Prediction layer： Forecast score .
   Insert picture description here
   

5. Loss function ：BPR Loss