Essays on machine learning for enhanced decision support in e-Commerce

Abstract

"E-commerce is proliferating, consequently, the need to make informed decisions. Many ecommerce platforms have adopted data-driven solutions in various applications, such as product recommendation and fraud detection. This dissertation explores how platform decisions on review recommendation, packaging design, and order delivery can be enhanced by leveraging advanced machine learning and optimization methods. The first essay addresses the review subset selection problem. Sifting through a large corpus of reviews is taxing on customers. Though prior works address this problem by recommending a subset of informative reviews, they fail to account for review quality resulting in suboptimal recommendations. In this essay, we propose two novel criteria to account for review quality and present an optimization framework to leverage the same in recommending high-quality, informative reviews. We establish the superiority of the approach by conducting user experiments on the Prolific platform. The second essay addresses the e-commerce box-sizing problem. E-commerce packages are notorious for their inefficient usage of space. More than one-quarter volume of a typical package comprises air and filler material. The inefficient usage of space results in increased transportation and material costs. Therefore, designing an optimal set of packaging box sizes is crucial. Prior approaches for solving this problem are not scalable. This essay proposes a scalable three-stage optimization framework combining unsupervised learning, reinforcement learning, and tree search to design optimal box sizes. We demonstrate the superiority of the approach by conducting experiments on real-world and synthetic datasets. The third essay addresses the last-mile delivery problem in e-commerce. The problem is crucial as leading e-commerce firms commonly report delivery failure rates of around 15%. Current approaches to the problem do not account for customer availability; therefore, schedules followed by delivery executives are optimized for the shortest tour distance resulting in missed deliveries. Missed deliveries cause a significant increase in fuel, storage, and package handling costs. We propose a prescriptive analytics framework for addressing this problem. Our framework generates delivery schedules by predicting appropriate time periods for order delivery. We demonstrate the benefits of our approach by performing numerical experiments on real-life delivery hub data obtained from a leading e-commerce firm."