Over the last two decades, digitization has been drastically shifting the way businesses operate and has provided access to high volume, variety, velocity, and veracity data. Naturally, access to such granular data has opened a wider range of possibilities than previously available. We leverage such data to develop application-driven models in order to evaluate current systems and make better decisions. We explore three application areas.
In Chapter 1, we develop models and algorithms to optimize portfolios in daily fantasy sports (DFS). We use opponent-level data to predict behavior of fantasy players via a Dirichlet-multinomial process, and our predictions feed into a novel portfolio construction model. The model is solved via a sequence of binary quadratic programs, motivated by its connection to outperforming stochastic benchmarks, the submodularity of the objective function, and the theory of order statistics. In addition to providing theoretical guarantees, we demonstrate the value of our framework by participating in DFS contests.
In Chapter 2, we develop an axiomatic framework for attribution in online advertising, i.e., assessing the contribution of individual ads to product purchase. Leveraging a user-level dataset, we propose a Markovian model to explain user behavior as a function of the ads she is exposed to. We use our model to illustrate limitations of existing heuristics and propose an original framework for attribution, which is motivated by causality and game theory. Furthermore, we establish that our framework coincides with an adjusted ``unique-uniform'' attribution scheme. This scheme is efficiently implementable and can be interpreted as a correction to the commonly used uniform attribution scheme. We supplement our theory with numerics using a real-world large-scale dataset.
In Chapter 3, we propose a decision-making algorithm for personalized sequential marketing. As in attribution, using a user-level dataset, we propose a state-based model to capture user behavior as a function of the ad interventions. In contrast with existing approaches that model only the myopic value of an intervention, we also model the long-run value. The objective of the firm is to maximize the probability of purchase and a key challenge it faces is the lack of understanding of the state-specific effects of interventions. We propose a model-free learning algorithm for decision-making in such a setting. Our algorithm inherits the simplicity of Thompson sampling for a multi-armed bandit setting and we prove its asymptotic optimality. We supplement our theory with numerics on an email marketing dataset.
Identifer | oai:union.ndltd.org:columbia.edu/oai:academiccommons.columbia.edu:10.7916/d8-3cw3-c550 |
Date | January 2020 |
Creators | Singal, Raghav |
Source Sets | Columbia University |
Language | English |
Detected Language | English |
Type | Theses |
Page generated in 0.0017 seconds