Machine learning practitioners often use confusion matrices, which compare predicted to actual class labels, to evaluate models. However, conventional confusion matrices do not support hierarchical and multi-output labels or other common, more complex data structures. As such, Görtler et al. develop Neo, a visual analytics system based on a new algebra that models confusion matrices as probability distributions. Neo enables ML practitioners to review additional metrics, carefully inspect model confusions, visualize hierarchical and multi-output labels, and share confusion matrix configurations with peers.