Dynamic Networks: Learning and Decision Making

Several daily phenomena around us can be modeled as time-evolving networks. Working with expressive and tractable models for the evolution of such networks can improve different prediction and decision-making tasks. While the literature has studied many approaches to model such networked phenomena partially, multiple gaps remain. This thesis is an effort to propose novel and scalable models and methods that capture temporal and spatial aspects of graph-structured data. We study three problems that involve dynamic networks and introduce new approaches to learn/optimize them. First, we focus on community detection and link prediction tasks in dynamic graphs. We propose efficient statistical inference algorithms for graph sequences by extending models and methods available for Stochastic Block Models. Next, we explicitly model the network effect in pricing products in a multi-stage decision-making setting. Under different scenarios, we develop pricing strategies that exploit social learning mechanisms to increase product adoption and firm revenue. Finally, we incorporate a knowledge graph that encodes item relationships into a temporal recommendation architecture via a suitable attention mechanism. Our approach improves the learning of user and item representations, especially for sparse datasets.