The focus of this thesis is on developing probabilistic models for data observed over temporal and graph domains, and the corresponding variational inference algorithms. In many real-world phenomena, sequential data points that are observed closer in time often exhibit higher degrees of dependency. Similarly, data points observed over a graph domain (e.g., user interests in a social network) may exhibit higher dependencies with lower degrees of separation over the graph. Furthermore, the connectivity structures that define the graph domain can also evolve temporally (i.e., temporal networks) and exhibit dependencies over time. The data sets observed over temporal and graph domains often (but not always) violate the independent and identically distributed (i.i.d.) assumption made by many mathematical models. The works presented in this dissertation address various challenges in modelling data sets that exhibit dependencies over temporal and graph domains. In Chapter 3, I present a stochastic variational inference algorithm that enables factorial hidden Markov models for sequential data to scale up to extremely long sequences. In Chapter 4, I propose a simple but powerful Gaussian process model that captures the dependencies of data points observed on a graph domain, and demonstrate its viability in graph-based semi-supervised learning problems. In Chapter 5, I present a dynamical model for graphs that captures the temporal evolution of the connectivity structures as well as the sparse connectivity structures often observed in temporal real network data sets. Finally, I summarise the contributions of the thesis and propose several directions for future works that can build on the proposed methods in Chapter 6.