Search

We model the work of a front-line service worker as a queueing system. The server interacts with customers in a multi-stage process with random durations. Some stages require an interaction between server and customer, while other stages are performed by the customer as a self-service task or with the help of another resource. Random arrivals by customers at the beginning and during an encounter create random lengths of idle time in the work of the server (breaks and interludes respectively). The server considers treatment of an infinite amount of back-office tasks, or tasks that do not require interaction with the customer, during these idle times. We consider an optimal control problem for the server's work. The main question we explore is whether to use the interludes in service encounters for treating back-office, when the latter incur switching times. Under certain operating environments, working on back-office during interludes is shown to be valuable. Switching times play a critical role in the optimal control of the server's work, at times leading the server to prefer remaining idle during breaks and interludes, instead of working on back-office, and at others to continue back-office in the presence of waiting customers. The optimal policy for use of the interludes is one with multiple thresholds depending on both the customers queueing for service, and the ones who are in-service. We illustrate that in settings with multiple interludes in an encounter, if at all, the back-office work should be concentrated on fewer, longer and later interludes., NA

Search engines and recommender systems have an important influence on the recruitment process, both for recruiters and job seekers. For both, they determine which job seekers/vacancies are recommended in what order, on résumé databases or recruitment websites. In doing so, they influence which vacancies job seeker apply to, or which job applicants are selected for an interview. To determine which vacancies/candidates to show and in what order, the search engine/recommender system uses estimations of the relevance of a vacancy/candidate for a job seeker/recruiter. To estimate this relevancy, the search engine/recommender system often only have partial information. In particular, only the search queries are known, and the click interaction on the search result. This interaction data shows which vacancies/candidates have been clicked on by users of the search engine/recommender system. This dissertation considers multiple algorithms, with the goal of improving estimations of relevance in the context of job/candidate recommender systems and search engines.

In this dissertation, I study how machine learning can be used to solve prominent problems in operations management and digital marketing. The primary motivation is to show that the application of machine learning can solve problems in ways that existing approaches cannot. In its entirety, this dissertation is a study of four problems—two in operations management and two in digital marketing—and develops solutions to these problems via data-driven approaches by leveraging machine learning. These four problems are distinct, and are presented in the form of individual self-containing essays. Each essay is the result of collaborations with industry partners and is of academic and practical importance. In some cases, the solutions presented in this dissertation outperform existing state-of-the-art methods, and in other cases, it presents a solution when no reasonable alternatives are available. The problems are: consumer debt collection (Chapter 3), contact center staffing and scheduling (Chapter 4), digital marketing attribution (Chapter 5), and probabilistic device matching (Chapters 6 and 7). An introduction of the thesis is presented in Chapter 1 and some basic machine learning concepts are described in Chapter 2.