Date of Award
August 2024
Degree Type
Dissertation
Degree Name
Doctor of Philosophy
Department
Management Science
First Advisor
Kaan Kuzu
Committee Members
Ehsan Soofi, Layth Alwan, Refik Soyer, Kaan Kuzu
Keywords
bayesian, dirichlet process, inventory management, recommender systems, semi parametric model, stochastic frontier model
Abstract
In today's competitive landscape, efficient production and effective product recommendationsystems are crucial for business success. Currently, there is a lack of comprehensive performance measures to evaluate and rank production lines based on their efficiency in turning resources into production outputs. Furthermore, firms need recommender systems that incorporate business objectives when ranking products on e-commerce channels. One objective that has not been explored enough in the literature is considering inventory management within recommender systems. In this thesis, we propose two novel models, one for assessing production line efficiency and another for enhancing product recommendation systems by considering inventory limitations, to address and fill these gaps. Historically, absolute and partial productivity measures have been commonly employed to compare various production lines within a firm. In the second chapter of this thesis, we develop a firm-specific benchmarking system by utilizing efficiencies estimated by Byesian stochastic frontier modeling as a production line performance measure. Our Bayesian stochastic frontier model estimates the production frontier, defined as the maximum possible output of the lines given the available resources. The closer a production line's observed output is to this frontier, the higher its efficiency is. The proposed model aims to modify the production frontier to account for the heterogeneity of products across various production lines and integrate existing productivity and econometrics literature to introduce a new structure for the technical inefficiency factor. The proposed structure is dependent on time and line-specific variables. We develop a Bayesian semi-parametric model with time-varying technical efficiency and assume a Dirichlet Process prior for the base technical inefficiency factor to circumvent the limitations of parametric prior distributions. This approach to line comparison enables the consideration of exogenous variables in estimating efficiency, incorporates stochasticity in performance analysis, and provides potential explanations for the underlying causes of low performance. Adopting a Bayesian framework, we obtain the posterior distribution of line efficiencies and propose using the stochastic ordering of technical efficiencies to cluster and rank groups of production lines. We first apply our model to a dataset provided by a major manufacturing firm. To extend our model beyond production line comparison, we also implement our model on a dataset comparing countries' efficiencies in life expectancies. Additionally, we compare the performance of our model with an alternative model from the literature and demonstrate that the proposed model yields more distinct groups of lines in terms of rankings. Finally, we highlight the managerial insights gained from clustering and stochastic ranking of production lines. In the second chapter, we pivot to address a critical gap in the development of recommender systems by integrating inventory management into traditional single-objective recommender systems. While traditional recommender systems have focused on maximizing customer engagement through user engagement metrics such as clicks and purchases, this narrow focus often fails to consider broader business objectives such as profit margins and inventory management. The recent challenges posed by the global pandemic, which led to raw material shortages and product scarcity, have underscored the drawbacks of traditional engagement-based recommenders which can boost demand for products that lack inventory and lead to unsatisfied customers. Here, we propose a novel recommender system that balances two potentially conflicting objectives by incorporating inventory management as a second objective into the traditional single objective user-centered recommender system design. By introducing and evaluating an inventory-aware recommender system, this study aims to enhance the practical applicability of recommender systems, ensuring they better serve both consumer interests and business goals. By integrating these perspectives, this thesis contributes to advancing both efficiency evaluation and recommender system design, offering practical solutions to improve operational performance as well as user and business satisfaction in modern business environments.
Recommended Citation
Nouri, Jessie, "Ranking Models for Product Manufacturing and Recommendation" (2024). Theses and Dissertations. 3606.
https://dc.uwm.edu/etd/3606