ICDM 2022: Risk Commodities Detection on Large-Scale E-Commence Graphs

[Slides]

This is the 3rd place solution for ICDM 2022: Risk Commodities Detection on Large-Scale E-Commence Graphs Competition.

Introduction

Recently, graph computing technologies, especially graph neural networks, have gained rapid development as well as wide application.

In the risk commodity detection scenario on e-commerce platforms, there is a fierce confrontation between attackers and risk detection systems. Malicious users always deliberately disguise risk information in order to avoid platform detection. By introducing graph data, we can alleviate the performance decline caused by the confrontation of malicious users.

In practice, the performance of the graph algorithm is often closely related to the quality of the graph structure. Malicious users usually forge their devices, and addresses to circumvent our detection. How to fully mine risk information in the graph structure data with a lot of noise is a very challenging problem. In addition, heterogeneous graph structure, as well as huge graph scale, are also challenges.

The competition provides a risk commodity detection dataset extracted from real-world risk scenarios at Alibaba. It requires participants to detect risky products using graph algorithms in a large-scale and heterogeneous graph with imbalanced samples.

Requirements

Higher versions should be also available.

• Python 3.7.3
• numpy==1.18.1
• torch==1.12.1+cu102
• torch-cluster==1.6.0
• torch_geometric>=2.1.0
• torch-scatter==2.0.9
• torch-sparse==0.6.14
• torch-spline-conv==1.2.1
• CUDA 10.2
• CUDNN 7.6.0

HeteroGNN

Model Overview

Data preparation

• Creat a folder /data in your workspace
• Download Session I and II datasets from here and put them in the data/session/ and data/session2, respectively.
• Extract zipped graphs: unzip icdm2022_session1_train.zip or unzip icdm2022_session2.zip
• Pre-process datasets (see follows)

Preprocessing

• Session I

cd code/
python process_data.py --session 1

Then you will get icdm2022_session1.pt in the /data folder.

• Session II

cd code/
python process_data.py --session 2

Then you will get icdm2022_session2.pt in the /data folder.

NOTE: other unnecessary files are manually removed after pre-processing.

How to Run the Model

Codes of Session I and Session II are marked by their suffixes _sess1 and _sess2.

Session I

cd code/
python main_sess1.py --lp --full --n-epoch 30  && python inference_sess1.py --lp 

• Running time ~20min
• main_sess1.py is to train the model and inference_sess1.py is to load the model and do inference
• --lp indicates using Masked Label Propagation during training and inference
• --full indicates using full training datasets (including validation set)
• We run 30 epochs in Session I to avoid overfitting

After finished, there would be two files in the code/ directly:

• model_sess1.pth: trained model
• session1_record.txt: running logs

Session II

cd code/
python main_sess2.py --n-epoch 100 --full && python inference_sess2.py

• Running time ~40min
• --full indicates using full training datasets (including validation set)
• Masked Label Propagation is not used in Session II as there are no labels provided

After finished, there would be two files in the code/ directly:

• model_sess2.pth: trained model
• session2_record.txt: running logs

File Structures

ICDM2022_competition_3rd_place_solution
├── code
│   ├── inference_sess1.py
│   ├── inference_sess2.py
│   ├── logger.py
│   ├── main_sess1.py
│   ├── main_sess2.py
│   ├── metapath.py
│   ├── model_sess1.py
│   ├── model_sess2.py
│   ├── process_data.py
├── data
│   ├── session1
│   │   ├── icdm2022_session1.pt
│   │   ├── icdm2022_session1_test_ids.txt
│   │   └── Readme.md
│   └── session2
│       ├── icdm2022_session2.pt
│       ├── icdm2022_session2_test_ids.txt
├── README.md
└── submit