[paper reading] kgat: knowledge graph attention network for recommendation

original text ：https://arxiv.org/pdf/1905.07854.pdf
Code ：https://github.com/xiangwang1223/knowledge_graph_attention_network

What: To provide more accuracy 、 Diverse and interpretable suggestions , Must go beyond the user - Modeling of object interaction , And consider Additional information （ such as KG）. In this paper , We studied Knowledge map (KG) The utility of , It does this by associating items with their properties , To break the Independent interaction hypothesis .（ By way of KG and user-item graph Combine , Capture high-dimensional connections ）【 High dimensional connection ：which connect two items with one or multiple linked attributes】

Why: Traditional approach , Such as factoring machine (FM) Think of it as a supervised learning problem , It assumes that each interaction is an independent instance , And encode the edge information . Because the relationship between instances or objects is ignored ( for example , The director of one film is also the actor of another film ), These methods are not enough to extract from the collective behavior of users Cooperative signal .（collaborative signal）

How: KGAT Recursively from the node's neighbors ( It can be a user 、 Item or attribute ) Propagate embed and aggregate , To refine the embedding of nodes , And use the attention mechanism to distinguish the importance of neighbors .

Result
Conclusion

This article emphasizes the importance of collaboration signals , And it aims to surpass users - Modeling object interaction , So consider adding additional information ; By way of 【 Knowledge map 】 and 【 user - Product interaction diagram 】 Combine , obtain 【 Collaborative knowledge map CKG】, To capture high-dimensional connections 、 Cooperative signal . The way to do it is , Through multi-layer embedding aggregation including attention mechanism, it comes from a node neighbor （ It can be user、item、attribute）, To update 、 Optimize embedded representation .

Model

The author begins with an example , Explain adding KG Then capture High dimensional connection 、 Cooperative signal Importance .
Collaborative filtering method , Focus on finding similar users （ Interact with a item）, So by i1,u1 Can find u4 and u5; Monitor learning methods , Emphasize looking for similar items （ Have the same properties ）, So by e1,i1 find i2.
The above method cannot be found u1 and u2、u3 of , Unable to find ,u1 and i3、i4 of .

1. CKG Embedding Layer

Energy fraction of triples （TransR）：（ The lower the score, the higher the confidence 【 There is more relationship between two entities 】）
among ,

Optimize through pairwise sorting losses ：

among ,
By randomly replacing real relationships , Training .（ Equivalent to negative sampling ）

【 The purpose of this layer is to learn optimization Entity 、 Embedded representation of relationships 】 It is better than direct random embedded representation .

2. Attentive Embedding Propagation Layers

This layer is based on GCN Recursive transfer embedded representation ; according to GAT Get the embedded representation of the joint level , Emphasize the importance of this high-dimensional connection .
Each layer contains 3 A component ：information propagation, knowledge-aware attention, and information aggregation（ Information dissemination 、 Knowledge perception, attention mechanism and information aggregation ）

Information dissemination

eh Self aggregation （ego-network）：
Spread information before entities and their neighbors .

Aggregate entities with their neighbors , Through the linear attenuation coefficient .

Knowledge perception attention mechanism

The attenuation coefficient is ：

Make in relation r In the space of , The fraction of the attenuation coefficient obtained , From entities h And entities t The relationship between , For the closer entities , The larger the attenuation coefficient obtained , stay 【 Information dissemination 】 in , The more information you keep .
For the sake of simplicity , Use the inner product to get the fraction of the attenuation coefficient .

Re pass softmax Normalize all scores ：

From this we can get , Capturing cooperative signals 、 When aggregating entity neighbors , Which entities are assigned higher scores .

In the above formula , There is more than just entity representation , There is also relationship . In order to get more information in communication , Not only the node representation as input , Also get entities h And entities t Relationship representation between .

Information aggregation

This floor is a 【eh】 and 【eh Self aggregation （ego-network）】 polymerization , Get a new entity h It means .
There are many ways ：

High dimensional propagation

The above process is a complete process of one-layer aggregation , You can stack multiple levels of results , Capture high-dimensional connections and higher-order neighbor ：

Each level eh Express , adopt 【 On the upper floor eh Express 】 and 【eh Self expression 】 obtain .

3. Prediction Layer

Go through the above steps , You can get user representation and item representation after multi-layer aggregation , Connect multiple representations to form a user representation and an item representation ：

thus , It enriches user representation and item representation .

Reference resources ：

1. KGAT：Knowledge Graph Attention Network for Recommendation：https://hikg.net/archives/123/