Online reputation systems are certainly the most overlooked 'heroes' of today's social Web. While these mechanisms are a vital element of every online transaction, they have received less consideration than some of their more well-known cousins, such as recommender systems or social networks, whose success would often not have been possible and tenable without their discrete but active backing. It then follows that despite their value and importance, the implementation of current reputation mechanisms has mostly been the result of trial-and-error. Resting on an economic perspective, this thesis regroups three chapters whose frameworks and findings aim at helping mechanism designers and researchers understand key mechanisms at play and develop more efficient online reputation systems. The first chapter examines the optimal number of ratings a reputation mechanism must make publicly available within an online marketplace in order to minimize cheating and maximize Pareto efficiency. I develop a moral hazard stage game featuring fictitious players which has the compelling property to prevent reputation effects from disappearing in the long run. I show that the number of ratings displayed by a reputation system is a fundamental predictor of market efficiency, and that the latter number should be kept minimal in order to maximize social welfare in the market – especially for economies proposing interactions with a high profit margin. The second chapter studies how different classes of reporting behaviours commonly found online affect the reliability of a reputation mechanism. I develop an iterative stochastic approximation model which I use to construct a behavioural measure of efficiency, so-called 'reporting bias'. I demonstrate that reporting bias tends towards its maximum when raters comply with the reports left by their predecessors. Following this result, I recommend to keep the rating interface separated from the rest of the reputation system. I also find that fake ratings are particularly harmful when one type of behaviour is present in the economy and suggest to counterbalance sybil attacks by displaying pairs of contrasted ratings. Finally, I defend the use of the arithmetic mean against the median as a way to compute reputation scores. The third chapter analyses how 5-star rating scales can lead to the formation of bimodal distributions of ratings within online marketplaces. Using a 2-time period model featuring altruistic raters, I identify the existence of a 'blind spot' of unrated transactions whose magnitude increases in the cost of rating and decreases in the number of buyers inhabiting the economy. Developing an additional model featuring Bayesian agents suffering from confirmatory bias, I show that non-binary rating scales can leave space to ambiguity and possibly wrong posteriors, even in the long run. Overall, results of the chapter hint that fine-grained rating scales best suit signalling reputation systems while coarse-grained scales should be preferred for sanctioning mechanisms.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:647645 |
| Date | January 2014 |
| Creators | Bersier, Florian |
| Contributors | Taylor, Greg; Vulkan, Nir |
| Publisher | University of Oxford |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
| Source | http://ora.ox.ac.uk/objects/uuid:deb557af-e848-481c-947e-94e0a4014994 |
Page generated in 0.0023 seconds