|Amazon Uses Machine Learning For Fraud Detection|
|Written by Sue Gee|
|Monday, 29 June 2015|
In an attempt to improve the quality of reviews on its US site, Amazon has introduced a new machine-learning platform that has been developed in-house.
Customer reviews and its 5-star rating system have been an important part of Amazon's success as an online marketplace so the prevalence of fake reviews on its site is a real problem for Amazon.
In April Amazon sued a company that openly traded in bogus 5-star reviews claiming:
"While small in number, these reviews threaten to undermine the trust that customers, and the vast majority of sellers and manufacturers, place in Amazon, thereby tarnishing Amazon's brand,"
This probably underestimates the scale of the problem and recently the UK Competition and Markets Authority opened an investigation into the problem of what it says are millions of fake online reviews.
Amazon already uses machine learning for its recommendation system, whereby when you make a purchase you are prompted to make further purchases based not only on what you have bought but also the buying habits and other customers with similar interests. Now it is extending the use of machine learning to improve the star rating system and the trustworthiness of reviews:
Amazon spokesperson Julie Law told CNet:
"The system will learn what reviews are most helpful to customers...and it improves over time. It's all meant to make customer reviews more useful."
The new system will give more weight to newer reviews, reviews from verified Amazon purchasers and those that more customers vote up as being helpful. A product's rating, which previously was a simple average of all reviews, will also become weighted using the same criteria, and so may change more often.
Initially this system is only being used in the United States and it doesn't really seem to go far enough in eliminating fraudulent reviews, something that AI would seem eminently suited to tackling by looking for outliers in a set of reviews, which if they relate to the same product, should be characterized by consensus. It would also seem to be just right for spotting reviewers who behave in non-typical ways. For example, a reviewer with one review of a highly technical book on programming in Lisp and lots of other reviews on non-programming topics.
At the end of the day, however, the problem is the basic motivation for customer reviews. Most customers simply want to buy a product and get on with their life. If the product turns out to be bad they might return to the review page and vent their anger, so biasing true reviews to the negative. Positive genuine reviews are most likely going to be rare because customers have no real reason to waste time submitting their thoughts. This allows a small number of paid reviewers to swamp the system.
With this in mind the simplest solution to fake reviews isn't AI but economics. To improve the quality of reviews what is needed is a good reason to write a review that is uncorrelated with the quality of the product. In this way the greater number of genuine customer reviews would swamp the smaller number of fraudulent ones. Perhaps Amazon should simply credit customer's accounts in return for a review - then they would all be paid-for reviews.
2023 Hour Of Code Announced
New activities for this year's Hour of Code have been announced. The focus for 2023,"Creativity with AI", is designed to let students learn about computer science and artificial intelligence.
How High Can Kotlin Go?
Having entered the Top 20 of the TIOBE Index in September 2023, Kotlin has continued its upward trend. Is it going to break into the Top 10 any time soon? Could it emulate Python and rise to the top?[ ... ]
or email your comment to: email@example.com