Product reviews are a great thing.
Fake reviews suck.
In the most recent example, an employee solicited paid reviews for his company’s products on Amazon’s Mechanical Turk – got to appreciate the progress.
How can you tell which reviews to trust? Trust is built out of relationship. You trust a site, a person, a brand, after your interactions accumulated enough positive history to earn that trust. With review sites, you may learn to trust a specific site, but that still doesn’t mean you trust a specific reviewer. I usually try to look at the reviewer’s history, and to look for the “human” side of them – spelling mistakes, topic changes, findings flaws and not just praising. But naturally, the adversary here is also informed, and will try to imitate these aspects…
Review sites attempt to bestow trust of their own on their members, to assist us. Amazon uses badges, and encourages users to provide their real name, using a credit card as the identity proof. Midrag is an Israeli service provider ratings site I recently used, that attaches identity to a cellular phone, with a login token sent over SMS. But when you want to attract a large number of reviews, you want to allow unvalidated identities too. Epinions, for example, builds a “web of trust” model based on reviewers trusting or blocking other reviewers. But with Epinions (and similarly Amazon) keeping their trust calculation formula secret, how can users be convinced that this metric fits their needs?
In reality, my model of trust may be quite different from yours. Two Italian researchers published a paper in AAAI-05 titled “Controversial Users demand Local Trust Metrics“, where they experimented with Epinions’ data on the task of predicting users’ trust score, based on existing trust statements. Their findings show that for some users, trust is not an average quantity, but a very individual one, and therefore requires local methods.
Trust metrics can be classified into global and local ones (Massa & Avesani 2004; Ziegler & Lausen 2004). Local trust metrics take into account the subjective opinions of the active user when predicting the trust she places in unknown users. For this reason, the trust score of a certain user can be different when predicted from the point of view of different users. Instead, global trust metrics compute a trust score that approximates how much the community as a whole trusts a specific user.
Have you spotted a familiar pattern?… Just exchange “trust” with “relevance”, and the paragraph will all of a sudden describe authority-based search (PageRank) versus socially-connected search (Delver). Local metrics were found to be more effective for ranking controversial users, meaning users that are assigned individual trust scores that highly deviate from their average score. The search equivalent can be considered queries that are for subjective information, where opinions may vary and an authority score may not be the best choice for each individual searcher.
To read more about trust metrics, see here: trustlet.org