Figure 1
Modeling approach. (A) Plate notation for the mixture of state space models. In this directed acyclic graph, each node represents a variable. Any two variables are independent after conditioning on their parents. Full shading denotes variables available during both training and testing, half shading denotes variables available during training but not necessarily during testing, and the remaining variables with no shading must be inferred from data. Plates indicate repetition. The plate on the left corresponds to the cluster-specific parameters for each of the \(n_c\) clusters. The larger plate governs the trajectory of each of the \(n_d\) samples. (B) Generative modeling process. The model independently assigns each individual to a cluster and then generates a sequence of states with corresponding observations according to the cluster-specific dynamics for the assigned cluster.
