Advances and Applications in Deep Learning

With the increasing computational power and massive amount of data, Deep Learning has attracted attention from both academic communities and industrial companies due to the improvements in prediction performance as compared to other conventional machine learning algorithms. It has almost become the industry standard for predictive analysis and has been applied to variable applications from computer vision to natural language processing, from medical science to automated driving. However, its application in business analytics, which refers to methods and practices that create value through data for individuals, firms, and organizations, has been limited, having a few numbers of research works utilizing it in this discipline. Therefore, this three-chapter rigorous and relevant research tries to connect the best of both worlds and contributes to both research fields. In the first chapter, we describe theoretical background of a neural network model which has been the foundation of Deep Learning and Artificial Intelligence and review the architecture variants, potential use cases, necessary requirements, and benefits. The second chapter establishes the sales elasticity of emotional displays in sales pitches by applying state-of-the-art Artificial Intelligence technologies to detect human faces and extract emotional displays from video data of 17,312 hours (2 years or 62 million frames) . So, we formulate new models to estimate the sales impact of emotional displays in the presence of marketing mix elements. Hereby, we can offer guidance and implications to firms on re-training of sales personnel to support the implementation of “selling with a straight face” as a maxim for their sales professionals, which is a provocative finding because it partially contradicts the external validity of the contagion theory. The last chapter tackles the problem of categorical variable with high cardinality which can exist quite a lot in business analytics cases. In this final study, we propose a simple and yet, effective approach to overcome the limitations of two dominant encoding schemes, i.e., one-hot encoding and target encoding, yielding competitive predictive performance, and substantial speed ups in training.