RE9: read the paper deal inductive link prediction for nodes having only attribute information

The gods were silent - personal CSDN Blog Directory

Title of thesis ：Inductive Link Prediction for Nodes Having Only Attribute Information
The paper ArXiv Download address ：https://arxiv.org/abs/2007.08053
The paper IJCAI Official download address ：https://www.ijcai.org/proceedings/2020/168（ There is a link to explain the video on this website , this TMD It's the first English paper video I've heard from a Chinese in so many days ！ Moved China , Finally, there is an English accent that I understand ！ I'm really fed up with maozi accent and third brother accent ！）（ The illustrations in this article are all taken from this video ）

This article is about 2020 year IJCAI The paper , Focus on inductive link prediction scene , New node only attribute, No structural information .
For the most part , The model proposed in this paper DEAL It's the difference attribute And graph structure （2 individual encoder, One is MLP, One is to directly take the unique heat code as the initialization matrix for node representation （ Even better than GCN The effect is good ）, They are all basic methods ）, Then use an alignment mechanism , Pass in training 2 individual encoder In addition to the loss function of optimization and contrast learning, a penalty term is added , Encourage the model to align the two representations of adjacent nodes （ Or similar to the loss function of comparative learning ）.
Use this to calculate the edge connection probability during the test ：

If it is inductive normal form , Don't count the first item .

It can be quoted DEAL Model LeSICiN¹ comparing ,LeSICiN Direct use MLP Calculate the probability score , Use this score to calculate the loss function . But the overall idea is similar .

↓ The following contents are basically not written , Just list a title , All to be filled .

1. Background & Motivation

2. DEAL Model

3. experiment

3.1 Data sets

3.2 baseline

3.3 Experimental setup

3.4 The main experimental results

This is not mentioned in the paper Cite. What model is it .

3.5 model analysis

4. Code reappearance

---

1. Re6： Reading papers LeSICiN: A Heterogeneous Graph-based Approach for Automatic Legal Statute Identification fro︎