Auction Fraud Detection Methods
I. User-Level Features
- Auction sites keep records of their users. For each user, we can divide the stored information into two parts: profile and past transactions. To determine a set of user-level features that distinguish perpetrators of fraud from honest users, we begin by learning from fraudulent cases that were widely publicized in newspaper articles and by examining the perpetrators involved. Our observations indicate that fraudsters tend to be short-lived. They exhibit burst-like trading patterns (many fake sales on a single day) and a bi-modal distribution of prices (cheap items sold to honest users and fictitious, expensive items sold to their alter-egos).
Download the Quick Reference Guide Here
II. Fraud Detection in Auction Listings
- This examines user-level features, meaning information intrinsic to individual users (e.g., “age” of the user, the number and prices of items sold/bought, the burst-like of the transaction times, etc.);
- This examines network-level features to detect suspicious patterns in the network of transactions between users, including trends (medians), fluctuations (standard deviations) and prices of items traded over time (first 15 days, 30 days, etc.).
III. Standard Deviation
- For example, one of the features is the standard deviation of prices of items sold within the first 15/30 days since the user registered. These features were previously evaluated to achieve a precision of 82% and a recall of 83%. The feature values can be extracted from the profiles and transaction history of users, available from the Web.
BELIEFS
A. We believe that the trends (medians)
B. Fluctuations (standard deviations)
C. Prices of items traded over time (first 15 days, first 30 days, etc)
STANDARD DEVIATION
1. For example, one of the features is the standard deviation of prices of items sold within the first 15/30 days since the user registered. These features were previously evaluated to achieve a precision of 82% and a recall of 83%.
2. The feature values can be extracted from the profiles and transaction history of users, available from the Web.
BELIEF PROPOGATION ALGORITHM The belief propagation is an algorithm used to infer the maximum likelihood state probabilities of nodes in a given graph. We use this technique to identify relationships not otherwise seen by the human eye. Here the red node indicates the hidden account that is interacting with all the other feedback in the seller’s history.
Post your comment on this topic.