How Groups Differ from Individuals in Learning from Experience: Evidence from a Contest Platform

We examine how groups differ from individuals in how they tackle two fundamental trade-offs in learning from experience-namely, between exploration and exploitation and between over- and undergeneralization from noisy data (which is also known as the "bias-variance" trade-off in the machine learning literature). Using data from an online contest platform (Kaggle) featuring groups and individuals competing on the same learning task, we found that groups, as expected, not only generate a larger aggregate of alternatives but also explore a more diverse range of these alternatives compared with individuals, even when accounting for the greater number of alternatives. However, we also discovered that this abundance of alternatives may make groups struggle more than individuals at generalizing the feedback they receive into a valid understanding of their task environment. Building on these findings, we theorize about the conditions under which groups may achieve better learning outcomes than individuals. Specifically, we propose a self-limiting nature to the group advantage in learning from experience; the group advantage in generating alternatives may result in potential disadvantages in the evaluation and selection of these alternatives.