[GNN report] Li Jia, Hong Kong University of science and technology: Rethinking graph anomaly detection - what kind of graph neural network do we need?

Catalog

1、 brief introduction

2、Rethinking Graph Neural Networks for Anomaly Detection

background

  learning model

Anomaly detection of existing drawings

​ edit   motivation

Background knowledge

  Method

  experiment

  summary

3、 Reference resources

---

1、 brief introduction

Guest speaker ： Li Jia （ Hong Kong University of science and technology ）

Report title ： Figure rethinking of anomaly detection — What kind of graph neural network do we need ？

Summary of the report ：

Figure neural network （GNN） It is widely used in anomaly detection of structured data , For example, malicious account detection on social networks 、 Financial transaction fraud detection, etc . For the first time, we analyze the possible impact of abnormal data from the perspective of spectral domain . The core findings are ： Abnormal data will lead to spectral energy “ Move right ” The phenomenon , That is, the spectrum energy distribution moves from low frequency to high frequency . Based on this discovery , We put forward again Beta Wavelet neural network （BWGNN）. It has several band-pass filters with local properties , Better capture “ Move right ” High frequency abnormal information generated . On four large-scale graph anomaly detection data sets ,BWGNN The performance of the model is better than that of the existing models .

About the reporter ：

Li Jia , Hong Kong University of science and technology Department of computer science and Hong Kong University of science and Technology ( Guangzhou ) Data science and analysis assistant professor ,2021 He graduated from the Chinese University of Hong Kong . Dr. Li Jia has many years of abnormal detection experience in industry , Worked for Google And Tencent . At present, its research mainly focuses on the anomaly detection of graph data , Reversible graph neural network and drug generation and medical health based on graph data .

reference

1.Rethinking Graph Neural Networks for Anomaly Detection

2.https://github.com/squareRoot3/Rethinking-Anomaly-Detection

2、Rethinking Graph Neural Networks for Anomaly Detection

background

 

 

  learning model

 

Anomaly detection of existing drawings

  motivation

  Start with finding the right filter

Background knowledge

  Graph Laplacian

Spectral decomposition （ External environment information ）

 ​​​​​

 

  Spectral clustering

 

  When the features become more and more , Low frequency energy is transferred to high frequency

Definition of low frequency energy

  Conclusion

When there is an anomaly in the graph or when the anomaly in the graph gradually becomes larger , The energy distribution after spectral decomposition is gradually transferred from the low-frequency part to the high-frequency part

  Verify the conclusion through hypothesis

If the low-frequency characteristic component becomes larger, the low-frequency energy becomes smaller

  Verify the conjecture ： On a synthetic dataset

  Verify the conjecture ： On real data sets

Verification on big data sets is difficult , Propose new concepts

verification

  Abnormal decrease is found , Low frequency energy rises （ The green dotted line is higher than the yellow dotted line ）

  Through the data value result verification

  Method

 

  It is found that wavelet transform is low-pass filtering

 

Discover wavelet GNN True coincidence , At the same time, determine to use beta kernal Of intuition

 

  go back to GNN On the design of , Still follow MPNN normal form  

  experiment

Framed is a big data set

 

 

  summary

3、 Reference resources

Video link || Li Jia, Hong Kong University of science and Technology ： Figure rethinking of anomaly detection — What kind of graph neural network do we need ？

LOGS The first 2022/07/16 period || Li Jia, Hong Kong University of science and Technology ： Figure rethinking of anomaly detection — What kind of graph neural network do we need ？_ Bili, Bili _bilibili