E-Trust

Now we move on to e-trust. In the last chapter we reviewed how people use technology to maintain and connect with more and more weak ties and peer-to-peer economy platforms build off the social networks to enable exchange among strangers. These exchange networks borrow a lot of features and lessons from eCommerce websites, such as reputation systems, and they share a lot of common properties with eCommerce sites, such as one needs to make a decision of whether to engage in a transction with a partner online in the lack of many information. So here we review literature on trust from eCommerce sites and review key elements of online trust, including reputation systems and their challenges.

Since the very beginning of eCommerce, trust was among the top concerns of system designers and researchers because of the significant amount of uncertain and risks involved. Sellers have an incentive to overstate the quality of the goods and omit crucial details of the flaws, hence charging a higher price. Buyers are concerned of possible frauds, and if they receive the goods first, they have the incentive of not paying the sellers back. Competitors of the sellers have an incentive to badmouth other sellers in order to get an advantage.

In fact, in the early days of eBay, the largest person-to-person online auction site in 2000, people were surprised that "the vast shuttling of both new and second hand goods among distant strangers" could happen at all given the considerable amount of risks [1]. The success of early online marketplace eBay was greatly attributed to its online reputation system. "After a transaction is complete, the buyer and seller have the opportunity to rate each other (1, 0, or 1) and leave comments (such as "good transaction," "nice person to do business with," "would highly recommend"). Participants have running totals of feedback points attached (visibly) to their screen names, which might be pseudonyms." [2]

Online reputation systems have since evolved to allow richer media, including photos or videos, but at its core, reputation systems centralize information from past exchange partners, and hence provide more information to potential exchange partners. According to Resnick et al. 2000, a reputation system must meet three challenges: "It must: (1) provide information that allows buyers to distinguish between trustworthy and non-trustworthy sellers (2) encourage sellers to be trustworthy, and (2) discourage participation from those who aren't." [2]. The key question therefore, is whether reputation system truthfully captures the information about past exchanges, and is beneficial to future exchange partners.

Over the years, many research showed that there are potential problems with reputation systems that may not capture information about past exchanges truthfully, including: positive bias (reputation inflation) and vulnerability to fake reviews.

Resnick et al studied one of the earliest and best known Internet reputation systems by eBay, using all transactions on eBay auction site from February through June 1999, and found that more than half of the transactions received feedback. Researchers emphasized that the reviews are highly positive, noting that "However, the 0.3% negative feedback rate on transactions (.6% of those that provided feedback) and 0.3% neutral feedback numbers from eBay, our principal data source, are highly suspicious." [1]. In addition, "there was a high correlation between buyer and seller feedback, suggesting that the players reciprocate and retaliate."

Similar positivity patterns were observed again and again. In a study conducted by Airbnb, non-reviewers were found to have worse experiences than reviews [3]. Perhaps it's simply hard to write negative reviews after you've spent the night at host's most intimate spaces -- their homes. Such selection bias in who writes online reviews created interesting phenomenon, summarized succinctly in the title of the paper "A first look at online reputation on Airbnb, where every stay is above average" [4], where the authors found that nearly 95% of Airbnb properties boast an average user-generated rating of either 4.5 or 5 stars (the maximum). The phenomenon of positive bias in reputation systems was framed as a game-theory equilibrium state, described by Horton and Golden in 2015 by conducting experiments on oDesk (rebranded to Upwork in 2015), a global freelancing online marketplace [5].

In addition to positive bias, reputation systems are also vulnerable to fake reviews. There is an ongoing race between our ability to detect fake reviews, and better ways to generate fake reviews. Back in 2011, simple linguistic features (n-grams and LIWC [6]) are sufficient to detect fake reviews generated by humans for hotels [7]. In 2017, with the latest development in deep learning, AI can generate fake Yelp reviews that are "not only evade human detection, but also score high on "usefulness" metrics by users" [8].

What is fascinating is that despite the problems, reputation systems still appear to work. A decade later, people expanded the realm of exchange from physical goods to resources, such as transportation and lodging. The rapid growing peer economy marketplaces, most prominently, Uber and Airbnb, created even bigger doubts around trusting potential exchange partners. Almost all these platforms make extensive use of reputation systems to "solve" the trust problem. A most recent experiment using Airbnb users as participants showed that reputation systems can increase trust between dissimilar users, hence adjusting the bias created by homophily (people's natural tendency to like others who are similar) [9].

Continued research is needed to fully understand the effect of reputation systems on the dynamics of online marketplaces. Cook et al. compiled a book titled "eTrust: Forming Relationships in the Online World: Forming Relationships in the Online World", summarizing dozens of crucial work on the effect of reputation systems on the dynamics of online marketplaces [10]. Broadly there are two methods, lab experiments, and field studies. For example, in chapter 3, Yamagishi et. al reproduced a lemons market in an experimental lab, and had participants to choose a level of quality for their product and the price they wish to sell for with the ability to change their own identities. It was found that the effect of reputation is smaller if sellers can change identities. In chapter 6, Snijders and Weessie studied the auction data from online programmer market and found that employers are much willing to pay a higher price for potential coders with high reputation (the reputation premium).

Finally, Kuwabara found in a 2015 study that recalling one's reputation on eBay made participants behave more trustworthily in the relevant roles (either as a buyer or as a seller) [11], indicating that their might be broader implications of reputation systems on behaviors in day-to-day life that we did not foresee before the reputation systems become so ubiquitous. When someone tweeted that one Uber drivers refused to pick him up as he as a rating lower than 4.7 (out of 5), it suddenly remind us with an uneasy feeling that Black Mirror episode, "Nosedive".

References