Online platforms such as recommendation systems and E-commerce websites have the capability to collect and store huge amount of data about users’ behaviors. One of the key design problems on those platforms is how users’ data is processed into useful information and displayed efficiently. This is complicated by the fact that users can be selfish and/or strategic but also can be error-prone. The former effect can lead to information cascades, in which users ignore their private information and blindly follow the other agents' action, leading to possibly a sub-optimal outcome.

In this talk, we present two models of observational learning among Bayesian agents on those platforms, where information cascades or herding can result. In the first model, information is suppressed by observation errors or noise. In the second model, additional information is introduced in the form of reviews. In both models, fundamentally non-intuitive results are discovered. We find out that neither a lower observation error level nor a higher review quality necessarily reduces the probability that agents make the wrong decision.