The ability to plan actions in order to achieve specific goals is essential for navigating daily life, from decisions about basic needs (e.g., food and medical care) to long-term objectives (e.g., financial planning, education, and preventive healthcare). Understanding how the healthy brain supports planning can help clarify how this capacity breaks down in a range of cognitive disorders, including age-related cognitive decline, Alzheimer’s disease, and addiction. In this talk, we present several new statistical machine learning methods designed to identify the algorithmic representations and structures underlying goal-directed decisions—the latent neural trajectories over time that give rise to particular choices. We begin by identifying patterns of ensemble neural activity over time using a latent representation learning approach. To address the challenge of integrating information across heterogeneous subjects, we introduce a new method, Integrated Latent Alignment (ILA), which combines deep representation learning with optimal transport (OT) theory. We then turn to the problem of characterizing structural connectivity among neurons to identify the activity patterns that generate these neural trajectories. To this end, we propose a new Graph Neural Network (GNN) model that leverages graph topology by representing brain structure as a complex network of interconnected units. The model consists of two components: an estimation component that captures latent relationships between neural features and trial outcomes, and an interpretation component that identifies an influential and compact subgraph by selecting a subset of node features that play a central role in the model’s predictions.