Integrating Advanced Machine Learning in Information Systems Research: What can Automated Machine Learning and Transfer Learning offer?

Abstract

Despite rapid advancements in supervised machine learning (SML) algorithms and their benefits for IS researchers, they have been slow in adopting SML in their research designs. In our literature review of information systems, we identify two principal limitations commonly faced by scholars: biases caused by selection (configuration and evaluation) of algorithms to fit a research task, and lack of sufficient labeled data to train SML algorithms. As a mitigating strategy, this article proposes automated machine learning and transfer learning. We conduct four case studies with both real-world and simulated data to show the effectiveness of proposed methods. The first three case studies demonstrate efficacy and ease of model building with automated machine learning framework. This become important to deal with common challenges in research settings of class-imbalance and high-dimensional data, where ensemble of multiple models is required for robust outcomes. The fourth case study demonstrate efficacy of transfer learning in extracting high-quality features from unstructured data even with less amount of labeled data, thus reducing the need to manually label huge quantities of data. These methods could help the community to develop a standardized format of presenting SML-based research, thus enhancing replicability of IS research.

Keywords: predictive analytics, automated machine learning, feature extraction, transfer learning, big data research, research guidelines