Development of Interpretable Machine-Learning Algorithms and Optimization Methods for Generation of Practical Insights
In recent years, we have been observing a revolution in the use of data science, machine-learning algorithms and optimization methods in “softer” areas such as human resources (HR), human behavior, mental illness, and learning disabilities, as well as in more conventional areas such as manufacturing and logistics systems. Despite these advances, there are still substantial gaps in our understanding of how users can implement machine-learning algorithms and optimization methods to address challenges in these domains. A user’s willingness to utilize the outputs of machine-learning algorithms is likely to be predicated upon his or her ability to understand the model’s behavior, rather than perceiving it as a black box. In this seminar we will present several research studies that are geared towards addressing the above challenges. In some of these studies, we develop new optimization methods and interpretable machine-learning algorithms, while in others, we adapt and implement existing algorithms, the results of which can provide practical insights into problems from various domains. More specifically, in this seminar we will present a variable-order Bayesian network (VOBN) algorithm and an ordinal classification tree based on a weighted information-gain measure, both of which are implemented in “softer” areas. We will also describe an optimal time-dependent hedging policy for production control under reputation-dependent demand.