A majority of the data transmitted and processed today has an inherent structured high-dimensional nature, either because of the process of encoding using high-dimensional codebooks for providing a systematic structure, or dependency of the data on a large number of agents or variables. As a result, many problem setups associated with transmission and processing of data have a structured high-dimensional aspect to them. This dissertation takes a look at two such problems, namely, communication over networks using network coding, and learning the structure of graphical representations like Markov networks using observed data, from an information-theoretic perspective. Such an approach yields intuition about good coding architectures as well as the limitations imposed by the high-dimensional framework. Th e dissertation studies the problem of network coding for networks having multiple transmission sessions, i.e., multiple users communicating with each other at the same time. The connection between such networks and the information-theoretic interference channel is examined, and the concept of interference alignment, derived from interference channel literature, is coupled with linear network coding to develop novel coding schemes off ering good guarantees on achievable throughput. In particular, two setups are analyzed – the first where each user requires data from only one user (multiple unicasts), and the second where each user requires data from potentially multiple users (multiple multicasts). It is demonstrated that one can achieve a rate equalling a signi ficant fraction of the maximal rate for each transmission session, provided certain constraints on the network topology are satisfi ed. Th e dissertation also analyzes the problem of learning the structure of Markov networks from observed samples – the learning problem is interpreted as a channel coding problem and its achievability and converse aspects are examined. A rate-distortion theoretic approach is taken for the converse aspect, and information-theoretic lower bounds on the number of samples, required for any algorithm to learn the Markov graph up to a pre-speci fied edit distance, are derived for ensembles of discrete and Gaussian Markov networks based on degree-bounded graphs. The problem of accurately learning the structure of discrete Markov networks, based on power-law graphs generated from the con figuration model, is also studied. The eff ect of power-law exponent value on the hardness of the learning problem is deduced from the converse aspect – it is shown that discrete Markov networks on power-law graphs with smaller exponent values require more number of samples to ensure accurate recovery of their underlying graphs for any learning algorithm. For the achievability aspect, an effi cient learning algorithm is designed for accurately reconstructing the structure of Ising model based on power-law graphs from the con figuration model; it is demonstrated that optimal number of samples su ffices for recovering the exact graph under certain constraints on the Ising model potential values. / text