The surprising outcomes of recent political events, such as the Brexit referendum and the latest US presidential election, are revealing how limited is our current understanding of social behavior in a highly interconnected world. A key challenge is to clarify how a massive information flow affects decision-making processes. As a contribution in addressing this issue, in this work we consider a social network where agents need to choose sequentially between two options, which could correspond to restaurants, brands or political parties. Decisions are based on two sources of information: a private signal that represents personal information, and social information given by the decisions of the previous agents. We consider rational agents that try to maximize an utility function that encodes their preferences. Additionally, we extend previous work [1,2] by considering that the agents might have incomplete knowledge of prior distributions and also that their preferences might be diverse, which introduces additional stochasticity that corresponds to ``social diversity''. To develop a deeper understanding we develop a communication theoretic interpretation of this scenario, treating it as a distributed decoding problem. From this point of view, a non-negligible social diversity is equivalent to additive noise and, hence, one would expect it to be detrimental for the collective learning process. Surprisingly, we prove mathematically and by numerical simulations that an adequate level of social diversity can help to mitigate unfavorable information cascades (c.f. [1]), which leads to improvements in the asymptotic learning performance. Social diversity however reduces the learning convergence speed and it benefits only affect latter agents, being hence undesirable for small networks.
