Data Mining to Detect Fraud in Auctions

[PhilipCohen]

The following article was published in “DSstar”, 27 June 2000
(Source: http://www.tgc.com/dsstar/00/0627/101834.html)


Using Data Mining to Detect Fraud in Auctions

by Ed Colet

In today’s Internet economy, auctions have become an increasingly common way to exchange goods and services. eBay is the market leader in the consumer-to-consumer space, and various companies have sprung up to apply the concept for business-to-business procurement. The nature of the bidding process in an auction seems to provide an intrinsic mechanism for both buyers and sellers to arrive at a fair price driven by the laws of supply and demand. But closer inspection reveals that the notion of a fair price can be fraudulently manipulated through the practice of “shill-bidding”. In this column, I discuss an example of this (on eBay), and point out how data mining technologies can be brought to bear to detect and prevent such fraudulent practices.

Shill bidding is the practice of a seller entering a bid on his/her own offering. Unlike an in-person auction, this is easy to do on the Internet because it is easy for a single person to have multiple Internet accounts and thus appear as separate persons. It is also easy for this concept to be extended from lone self-bidders to rings of shill bidders that conspire together with the intent of artificially creating high demand for the item and driving up the price.

As reported in the New York Times Technology section on June 2 2000, an analysis of eBay’s auction records suggested evidence of shill bidding. A painting whose opening bid was 25 cents had bidding prices that escalated to $135,805 before eBay suspended the seller and voided the sale. Records showed that there were 33 Internet names that repeatedly bid on one another’s offerings—and a subset of these names rarely placed a winning bid. The difficult question is whether the practice of shill bidding occurred, or if the overlap is a natural reflection of demand.

Although the average cost of an item on eBay is $40, one may think that the costs of shill bidding on items are negligible. But for the individual who bid $126,000 for the painting in question and was willing to raise another $125,000 [sic] to win the bid, the costs are not negligible. Consider also that auction sales in the US are expected to grow to $6.4 billion, and eBay having 90% of the market is expecting revenues of $400 million. In terms of transactions, there are 1000 bids a minute for 4.5 million items. Thus, both in terms of dollars and the amount of activity, auctions are a large scale activity. If fraud is widespread, costs are not negligible.

The large-scale nature of auction transactions can make it difficult to ferret out fraudulent practices using standard analytical methods. On eBay, bidding histories and user feedback records are not stored for long, and the progressive sequence of an individual’s bid is apparently not stored due to cost considerations and storage capacity limitations. It appears that whatever analysis eBay does conduct is on a limited data set and performed after the auction has been closed. A better approach would be to detect fraud dynamically (live) because sales need not be voided, and more data need not be stored if capacity and costs still remain an issue.

Fraud detection methods of data mining can be applied to this problem quite readily. Three important elements of a data mining application/solution are present. These are the ability to handle large amounts of data, suitable analytical methods and algorithms, and the availability of domain expertise. Data mining technologies are designed to look for subtle patterns in large data stores. The large amounts of data generated by auction transactions are thus not likely to be a problem. In terms of analytical methods, there are already numerous algorithms for detecting co-occurrences (of bids) in datasets. Determining whether co-occurring bids are fraudulent or if the overlap is legitimate can be decided in part by calculating the rate that these co-occurring bids turn out to be the winning bid (fraudulent bids are rarely the winning bid) as well as incorporating some domain knowledge. For example, it is known that user ids of rings of shill-bidders tend to belong to people that are geographically close to each other. Integrating such analytical methods, coupled with domain knowledge can result in a data mining application that can detect and minimize fraudulent practices in the auction arena.

________________________________________
Ed Colet is the Acting Director of Research at Virtual Gold Inc., responsible for developing analytical methods for data mining and for investigating human factors and usability issues of business intelligence systems. At present, he is in the final stage of completing a doctoral dissertation in the Cognition and Perception program at New York University’s Department of Psychology. Ed has also worked for IBM Research at the T.J. Watson Research Center. At IBM, Ed was a member of the group that developed Advanced Scout, the data mining application for NBA teams. His research interests focus on statistical methods and human factors.

For more information, see www.virtualgold.com.
_________________
Clearly, the lunatics at eBay have taken over the asylum and are bent on burning it down.
“The difference between genius and stupidity is that genius has its limits.” ~ Albert Einstein.

[Ebuster]

Good to see you again and i can see you are giving eBay a hard time.

Do you have anymore shill bidding leads for me to follow up ?

best regards

eBuster