Date of Award

August 2023

Degree Type


Degree Name

Doctor of Philosophy


Management Science

First Advisor

Huimin Zhao

Committee Members

Huimin Zhao, Yang Wang, Scott Schanke, Xiang Fang, Gang Chen


Artificial Intelligence, Business Analytics, Data Mining, Predictive Analytics


The big data era has provided researchers with challenges and opportunities for data-centric research. On the one hand, recent developments in AI technology have allowed advanced techniques to process text/image/audio/video and graph-structured data, providing new opportunities to employ big data for explanatory and predictive analytics in information systems research. On the other hand, the field requires a new level of artificial intelligence–transparent, robust, and ethical AI–to facilitate reliable business decision-making. My three dissertation essays apply, develop, and enhance state-of-the-art AI methods, leveraging various data sources as well as domain knowledge synthesis, to deal with issues in business and healthcare fields.

In Essay 1, I investigate the possibility of using deep learning models for Computed Tomography (CT) localizer image reconstruction. CT has become an important clinical imaging modality, as well as the leading source of radiation dose from medical imaging procedures. Modern CT exams are usually led by two quick orthogonal localization scans, which are used for patient positioning and diagnostic scan parameter definition. These two localization scans contribute to the patient dose but are not used for diagnosis purposes. I investigate the possibility of using deep learning models to reconstruct one localization scan image from the other, thus reducing the patient dose and simplifying the clinical workflow. I propose a modified encoder-decoder network and a scaled mixture loss function specifically for the focal task. Experiment results indicate that although the reconstructed abdominal CT localization images may lack some details on the internal organ structures, they could be used effectively for tube current modulation calculation and patient positioning purposes, leading to a reduction of radiation dose and scan time in clinical CT exams.

In Essay 2, I propose a robust meta-graph learning method for multimodal time series prediction. Multimodal time series prediction is a difficult problem given the intricate feature interrelationships. I explore interrelationships of multilevel features in multimodal time series data and disentangle the intricate interrelationships with a robust meta-graph learning method named RMGL. The design of RMGL is rooted in theoretical foundations regarding graph convolutional networks and a novel graph attention mechanism. The core of RMGL is a meta-graph composed of three hierarchically interconnected graphs, representing feature-wise, modality-wise, and time-step-wise interrelationships, respectively. The interconnections across the graphs allow feature representations to propagate simultaneously, thereby quantifying multilevel feature interrelationships with graph structures synchronously and efficiently. Furthermore, RMGL introduces a novel weight regularization scheme to effectively learn the meta-graph for prediction based on the low-pass nature of graph convolutional filters. RMGL outperformed state-of-the-art alternatives in an empirical evaluation with a financial risk prediction task. Ablation experiments and further analyses indicated the effectiveness of RMGL.

In Essay 3, I propose a knowledge-enhanced, transformer-based text categorization model to detect employee trust indices from employee reviews. The indices of Employee Trust Model (ETM) are intangibles. Extant measurement options that require members of an organization to complete surveys make it difficult to collect data from large samples of firms across times. The use of small samples has led to conflicting results in managerial and finance research and made findings less appealing to practitioners. Furthermore, the absence of data in the time dimension has restricted analytical methods in use and limited the application of theoretical frameworks. I propose DeepEmployee, a novel design artifact based on automated text classification, to detect ETM indices from employee-generated reviews. DeepEmployee stems from design science research and includes three cohesive and complementary parts: (1) domain-specific knowledge construction based on theoretical frameworks in the management field, (2) a state-of-the-art deep learning design artifact that incorporates domain-specific knowledge to improve performance, and (3) a rigorous two-part evaluation of improvements in ETM detection and increased explanatory and predictive power in downstream tasks.

Available for download on Thursday, August 29, 2024