In this paper, we present a principled probabilistic framework – GroupBox – for making recommendations to groups. GroupBox is able to model user influence within a group, the suitability of an item to a group context, and the differences in user preference between individual and group contexts. Efficient scalable inference algorithms are used for GroupBox, which makes it applicable to large-scale datasets. We run experiments on a large-scale TV viewing dataset collected by Nielsen and show how the model can be used to understand both context and influence. The experimental results on the large scale real data provide a deep understanding of the individual behaviours in group context.