GRAPH NEURAL NETWORKS

Srijith Rajamohan, Ph.d.

CONTENTS

1. Tools

2. Developments

 

Tools

Some popular tools for GNNs

DGL: Low-level but has integrations with a lot of frameworks like Pytorch and Keras. List of official examples here. Even has a large-scale knowledge graph library.  Includes a Pytorch example for graph explainability.

Pytorch Geometric: 

Stellargraph: High-level graph library, has plenty of built-in algorithms. List of examples here. Interpretability examples here.

Spektral: GNN library built on top of TensorFlow 2

Keras:  Some examples using just Keras

EVOLUtion of GNN

Graph Neural Network-based Fraud Detection from Research to Application

Yingtong Dou, University of Illinois at Chicago

Potential APPLICATIONS

  • Transaction fraud detection vs. fraudster detection
    • Anomaly detection - non-normal transactions based on historical patterns
    • Community detection - fraudsters share common entities and have clustering behavior
  • Recommendation engine for products
    • Link prediction in heterogeneous graphs
  • Knowledge graphs

FRAUD detection

TYPES OF RELEVANT FRAUD

Graph Neural Network-based Fraud Detection from Research to Application, Yingtong Dou, University of Illinois at Chicago

REVIEW of FRAUD

WHy graphs?

  • Better relational reasoning
  • Better generalization
  • Semi-supervised learning
  • Applications
    • GCNs to detect money laundering

Financial crimes

  • Payment fraud
    • Card Not Present (CNP)  transactions account for 70% of all payment fraud
    • Online transactions are vulnerable
    • Mobile payment channels are used for money laundering
  • Identity theft
    • Can take time to identify, 58% discovered after 4 months
    • Check fraud and credit card fraud, new account created fraud
    • Synthetic ids where real information is blended with manufactured information
  • Account takeover
  • Money laundering (about 1 trillion euros)
    • Placement of funds in small sums into the financial system
    • Layering - multiple transfers between shell corporations
    • Integration - integrating the funds into the financial system

GRAPH use-cases

  • Transaction risk scoring
    • Happens in real-time
  • Entity scoring using monetary and non-monetary information
    • Black/gray list generation to avoid or monitor behavior
  • Sub-graph analysis for criminal organizations
  • Evolving graphs, streaming GNNs, temporal networks for dynamic graphs
  • Pattern extraction with attention networks
    • Identify history buildup for fraudulent transactions or account takeover

GEM (2018)

  • Landmark paper using heterogeneous graph data
  • Detect malicious accounts in Alipay
  • Code is here

Graph Neural Network-based Fraud Detection from Research to Application, Yingtong Dou, University of Illinois at Chicago

GAS (2019)

  • Detection spam reviews
  • Encode each heterogenous graph separately
  • Code is here

Graph Neural Network-based Fraud Detection from Research to Application, Yingtong Dou, University of Illinois at Chicago

CARE-GNN (2020)

  • Detection spam reviews and malicious reviewer
  • Addresses the class imbalance problem in fraud detection
  • Integrated with DGL
  • Code is here

Graph Neural Network-based Fraud Detection from Research to Application, Yingtong Dou, University of Illinois at Chicago

FD-NAG (2021)

  • Fraud detection in ride-sharing
  • Convert relationship data to a homogeneous graph by encoding edge features
  • Uses contrastive learning in fraud detection

Graph Neural Network-based Fraud Detection from Research to Application, Yingtong Dou, University of Illinois at Chicago

GRAPH architectures

Graph Neural Network-based Fraud Detection from Research to Application, Yingtong Dou, University of Illinois at Chicago

In progress...

dum_var = np.array([0.0,0.0]) 
res = numba_svd_derivative
(j, self.nrow, self.ncol, self.projection_components, \
      self.u, self.v, self.Xorig, self.X, d_sq, dum_var)
Derivative[:,:,:] = res

Pros:

The font creates a flow and can be likened to someone's actual handwriting: this creates a more personal effect to the magazine. No other magazine usually has this type of font, so people would be intrigued to check this magazine out.

cons:

Quite a "squashed" font. I could imagine that it doesn't take up the ideal space a masthead should. It lacks boldness and it could potentially not gather interest if the masthead isn't prominent enough on the front cover.

Anchorage

examples

PROS:

Something that gives the impression of contemporary or modern art. Very minimalist but effective in portraying modern typography. The font is also quite graphic and gives an illusion, helping give a unique appeal if used on a page.

cONS:

Hard to see in some areas, some people might find it a nuisance if they can't read the important article (which is usually promoted by the anchorage).

pros:

Uppercase letters proves effective in attracting attention; it is not highly distracting enough to divert attention from the masthead too. Very elegant and gives a timeless appeal if used on the front cover.

Cons:

Reminiscent to Vogue, which some readers might not like as it is usually associated with female audiences. The font is also quite formal, which might not suit my genre or music magazine.

main

text

example

Made with Slides.com