safe-graph / ugfraud

An Unsupervised Graph-based Toolbox for Fraud Detection

An Unsupervised Graph-based Toolbox for Fraud Detection

Introduction: UGFraud is an unsupervised graph-based fraud detection toolbox that integrates several state-of-the-art graph-based fraud detection algorithms. It can be applied to bipartite graphs (e.g., user-product graph), and it can estimate the suspiciousness of both nodes and edges. The implemented models can be found here.

The toolbox incorporates the Markov Random Field (MRF)-based algorithm, dense-block detection-based algorithm, and SVD-based algorithm. For MRF-based algorithms, the users only need the graph structure and the prior suspicious score of the nodes as the input. For other algorithms, the graph structure is the only input.

Meanwhile, we have a deep graph-based fraud detection toolbox which implements state-of-the-art graph neural network-based fraud detectors.

We welcome contributions on adding new fraud detectors and extending the features of the toolbox. Some of the planned features are listed in TODO list.

If you use the toolbox in your project, please cite the paper below and the algorithms you used :

@inproceedings{dou2020robust,
  title={Robust Spammer Detection by Nash Reinforcement Learning},
  author={Dou, Yingtong and Ma, Guixiang and Yu, Philip S and Xie, Sihong},
  booktitle={Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery \& Data Mining},
  year={2020}
}

Useful Resources

• PyGOD: A Python Library for Graph Outlier Detection (Anomaly Detection)
• DGFraud: A Deep Graph-based Fraud Detection Toolbox
• Graph-based Fraud Detection Paper List
• Awesome Fraud Detection Papers
• Attack and Defense Papers on Graph Data
• PyOD: A Python Toolbox for Scalable Outlier Detection (Anomaly Detection)
• PyODD: An End-to-end Outlier Detection System
• DGL: Deep Graph Library
• Outlier Detection DataSets (ODDS)

Table of Contents

• Installation
• User Guide
• Implemented Models
• Model Comparison
• TODO List
• How to Contribute

Installation

You can install UGFraud from pypi:

pip install UGFraud

or download and install from github:

git clone https://github.com/safe-graph/UGFraud.git
cd UGFraud
python setup.py install

Dataset

The demo data is not the intact data (rating and date information are missing). The rating information is only used in ZooBP demo. If you need the intact date to play demo, please email [email protected] to download the intact data from Yelp Spam Review Dataset. The metadata.gz file in /UGFraud/Yelp_Data/YelpChi includes:

• user_id: 38063 number of users
• product_id: 201 number of products
• rating: from 1.0 (low) to 5.0 (high)
• label: -1 is not spam, 1 is spam
• date: data creation time

User Guide

Running the example code

You can find the implemented models in /UGFraud/Demo/ directory. For example, you can run fBox using:

python eval_fBox.py 

Running on your datasets

Check out the data_to_network_graph function in /UGFraud/Demo/demo_pre.py to convert your data into the networkx graph.

In order to use your own data, you have to provide the following information at least:

• a dict of dict:

'user_id':{
        'product_id':
                {
                'label': 1
                }

• a dict of prior

You can use dict_to networkx(graph_dict) function from /Utils/helper.py file to convert your graph_dict into a networkx graph. For more details, please see data_to_network_graph.py.

The structure of code

The /UGFraud repository is organized as follows:

• Demo/ contains the implemented models and the corresponding example code;
• Detector/ contains the basic models;
• Yelp_Data/ contains the necessary dataset files;
• Utils/ contains the every help functions.

Implemented Models

Model	Paper	Venue	Reference
SpEagle	Collective Opinion Spam Detection: Bridging Review Networks and Metadata	KDD 2015	BibTex
GANG	GANG: Detecting Fraudulent Users in Online Social Networks via Guilt-by-Association on Directed Graph	ICDM 2017	BibTex
fBox	Spotting Suspicious Link Behavior with fBox: An Adversarial Perspective	ICDM 2014	BibTex
Fraudar	FRAUDAR: Bounding Graph Fraud in the Face of Camouflage	KDD 2016	BibTex
ZooBP	ZooBP: Belief Propagation for Heterogeneous Networks	VLDB 2017	BibTex
SVD	Singular value decomposition and least squares solutions	-	BibTex
Prior	Evaluating suspicioueness based on prior information	-	-

Model Comparison

Model	Application	Graph Type	Model Type
SpEagle	Review Spam	Tripartite	MRF
GANG	Social Sybil	Bipartite	MRF
fBox	Social Fraudster	Bipartite	SVD
Fraudar	Social Fraudster	Bipartite	Dense-block
ZooBP	E-commerce Fraud	Tripartite	MRF
SVD	Dimension Reduction	Bipartite	SVD

TODO List

• Homogeneous graph implementation

How to Contribute

You are welcomed to contribute to this open-source toolbox. Currently, you can create issues or send email to [email protected] for inquiry.

ugfraud's Issues

Scale function bug?

Is the else condition indented wrongly?

def scale_value(value_dict):
    """
    Calculate and return a dict of the value of input dict scaled to (0, 1)
    """

    ranked_dict = [(user, value_dict[user]) for user in value_dict.keys()]
    ranked_dict = sorted(ranked_dict, reverse=True, key=lambda x: x[1])

    up_max, up_mean, up_min = ranked_dict[0][1], ranked_dict[int(len(ranked_dict) / 2)][1], ranked_dict[-1][1]

    scale_dict = {}
    for i, p in value_dict.items():
        norm_value = (p - up_min) / (up_max - up_min)
        if norm_value == 0:  # avoid the 0
            scale_dict[i] = 0 + 1e-7
        elif norm_value == 1:  # avoid the 1
            scale_dict[i] = 1 - 1e-7
    else:
        scale_dict[i] = norm_value

    return scale_dict

cannot import name 'Detector' most likely due to a circular import

Performing a simple import as outlined in testing.py

import sys
import os
__file__ = "~/env/lib/python3.8/site-packages/UGFraud"
sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))
from UGFraud.Demo.eval_fBox import *

However, this produces the below error:

---------------------------------------------------------------------------
ImportError                               Traceback (most recent call last)
~/env/lib/python3.8/site-packages/UGFraud in <module>
      3 __file__ = "~/env/lib/python3.8/site-packages/UGFraud"
      4 sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))
----> 5 from UGFraud.Demo.eval_fBox import *

~/miniconda3/lib/python3.8/site-packages/UGFraud/__init__.py in <module>
      1 # -*- coding: utf-8 -*-
      2 
----> 3 from . import Detector
      4 from . import Utils
      5 

ImportError: cannot import name 'Detector' from partially initialized module 'UGFraud' (most likely due to a circular import) (~/miniconda3/lib/python3.8/site-packages/UGFraud/__init__.py)

m-zoom & d-cube

good job!! Whether there is a plan for m-zoom and d-cube algorithm？