| This thesis is organized into two distinct parts; however, both explore different facets of large-scale machine learning. The first chapter is concerned with methodological aspects of machine learning, namely the design of new distributed optimization algorithms, whereas the last two chapters consider applications of machine learning that lie at the interface of social network analysis and marketing.;The motivation for the first chapter is that in recent years, the size of data sets has exceeded the disk and memory capacities of a single computer. This has created the need to employ parallel and distributed computing in order to perform critical machine learning tasks. Given that optimization is one of the pillars of machine learning and predictive modeling, distributed optimization methods have recently garnered ample attention in the literature. Although previous research has mostly focused on settings in which either the observations, or features of the problem at hand are stored in distributed fashion, the situation where both are partitioned across the nodes of a computer cluster (doubly distributed) has not been studied extensively. We propose two doubly distributed optimization algorithms. The first falls under the umbrella of distributed dual coordinate ascent methods, while the second belongs to the class of stochastic gradient/coordinate descent hybrid methods. We conduct numerical experiments in Spark using real-world and simulated data sets and study the scaling properties of our methods. Our empirical evaluation of the proposed algorithms demonstrates the outperformance of a block distributed ADMM method, which, to the best of our knowledge is the only other existing doubly distributed optimization algorithm.;The second and third chapters address two fundamental problems in marketing: customer retention, and product adoption, respectively. In the second chapter, we explore the power of social networks in predicting non-contractual customer behavior. The last decade has seen a rapid emergence of non-subscription-based services. Predicting spending patterns in such settings is particularly challenging due to the capricious purchasing behavior that customers often exhibit. We study the extent to which knowledge of a customer's social network can enhance the accuracy of forecasting their behavior in terms of future: (1) activity, (2) transaction levels and (3) membership to the group of best customers. We conduct a dynamic analysis on a sample of approximately one million users from the most popular peer-to-peer (P2P) payment application, Venmo. Our models produce high quality forecasts and demonstrate that social networks lead to a significant boost in predictive performance primarily during the first month of a customer's lifetime, thus providing a remedy to the "cold-start" problem. Finally, we characterize significant structural differences with regard to network centrality, density and connectivity between the top 10% and bottom 90% of users immediately after joining the service. We discuss how these structural dissimilarities provide a path towards proactive marketing and improved customer acquisition efforts.;In the third chapter, using the same data set as in the second chapter, we investigate the structural aspects behind the adoption of Venmo. The most distinct quality of our data set is that Venmo transactions reflect offline activities combined with the scale of the online world. We use the framework of structural virality, which has primarily been utilized in the spread of online content, as a vehicle to investigate the diffusion of an offline product. Structural virality serves as a tool for differentiating between the two main forms of diffusion, namely a single broadcast that reaches directly a large number of people, and a viral mechanism that follows a more organic, P2P growth. We explore the nature of adopting Venmo from two perspectives: the traditional approach that considers adoption to be a binary event, and a holistic approach that takes into account the behavior of users in the post-adoption period. Our key finding indicates that not all adoptions are equal. More specifically, our results show that broadcast and viral mechanisms are positively correlated with the speed of diffusion and user engagement, respectively. In the face of this trade-off, we build a model to predict broadcast and viral structures, which can help stakeholders proactively devise network interventions that best suit their objectives. |
