We report the development and validation of a principled analytical approach to reveal the manner in which diverse mouse home cage behaviors are organized. We define and automate detection of two mutually-exclusive low-dimensional spatiotemporal units of behavior: "Active" and "Inactive" States. Analyses of these features using a large multimodal 16-strain behavioral dataset provide a series of novel insights into how feeding, drinking, and movement behaviors are coordinately expressed in Mus Musculus. Moreover, we find that patterns of Active State expression are exquisitely sensitive to strain, and classical supervised machine learning incorporating these features provides 99% cross-validated accuracy in genotyping animals using behavioral data alone. Altogether, these findings advance understanding of the organization of spontaneous behavior and provide a high-throughput phenotyping strategy with wide applicability to behavioral neuroscience and animal models of disease.