Literature Review, GraphRAD: A Graph based Risky Account Detection System - herculescw/graph-fraud-detection-papers GitHub Wiki

Background:

Fraudsters use stolen victim's account information to make purchases through Amazon online purchase platform.

Assumption:

Fraud accounts are tightly connected to each other inside the fraud communities.

Seeding and Outcome:

Give a few fraud accounts as the seeds, return fraud communities and communities cannot be too much overlapped and too big.

Scoring and Ranking:

Every detected fraud account is scored and ranked, and fraud experts will prioritize reviewing transactions by ranking.

System Architecture

Transaction Record, a dataset with attributes on transaction time, account id, shipping address, decision results(trusted, fraud, risky)

Graph Generator, construct the graph, link the accounts by some common sharing attribute, such as shipping address

Seeding, selected blacklist seeds, the data marked as a fraud in the dataset

Community Detection, local fraud community detection, propagated from blacklist seeds, using a personal PageRank ACL

Screen+Merge, filtering and merging, filter small risky communities, and merge left communities into a big-comm

Community Extractor, use hierarchical clustering to get the final community.

Feature Extraction, extract the big-comm nodes' features.

Scoring, semi-supervised model to score every node.

---

A Chinese version review on this paper, https://zhuanlan.zhihu.com/p/75123819