Inderscience
Publishers
Fake
reviews do nothing for the confidence of customers buying products and services
online, they also damage company reputations and can lead to ill feeling about
the online marketplace itself.
Now,
researchers in China have devised an algorithm to help weed out fake reviews on
ecommerce sites. They publish details this month in the International
Journal of Services Operations and Informatics.
Song Deng of the Jiangxi University of Finance and Economics, Nanchang, China, explains how our shopping habits have changed and more and more people are buying products and services online.
One of the mainstays of the modern sales
website are customer reviews and there are even complete sites that offer
consumers a place to discuss their experiences with a given product.
Over
the years, there have been several scandals regarding large numbers of fake
reviews on major online marketplaces and sites offering travel advice and
holiday packages.
There is an urgent need to develop a robust algorithm that
can detect the fakers and remove their hyperbole and give consumers a truer
picture of whether a given product is an five-star or a no-star item.
In other
words, we need an automatic lawnmower to cut down the "astroturfing,"
the artificial grass-roots marketing of products.
Deng's
method recognises deceptive reviews based on how the posters has behaved
previously and the content of their earlier reviews.
First, it builds a
recognition model that can spot fake reviewers, ghostwriters and paid members
of the "so-called "water army" based on the number of reviews,
frequency and length.
It then looks at content features, such as review length,
the degree of professionalism, the emotional density, the format and any
obvious biases.
Finally, the algorithm applies an unsupervised clustering
algorithm based on F statistics and a feature degree.
When
combined, these techniques outshine earlier detection algorithms for reviews of
cars, smart phones and computers.
Fundamentally, the system combines the
advantages of behavior feature and content feature recognition to improve
accuracy.