Machine learning has become a priority for organizations as they move towards data monetization strategies involving predictive analytics applied across the enterprise. Despite the potential for immense value, many aspirational firms’ predictive capabilities are impeded by data management constraints including data quality issues, privacy regulations, organizational silos to data integration, the need to monetize an array of structured and unstructured data types, all while dealing with increasingly complex and dynamic environments. Domain generalization is the branch of transfer-based machine learning designed to handle such situations where training data is unavailable in the application domain, and where the domain patterns might shift (known as training-serving skew). However, the d...[ Read more ]

Machine learning has become a priority for organizations as they move towards data monetization strategies involving predictive analytics applied across the enterprise. Despite the potential for immense value, many aspirational firms’ predictive capabilities are impeded by data management constraints including data quality issues, privacy regulations, organizational silos to data integration, the need to monetize an array of structured and unstructured data types, all while dealing with increasingly complex and dynamic environments. Domain generalization is the branch of transfer-based machine learning designed to handle such situations where training data is unavailable in the application domain, and where the domain patterns might shift (known as training-serving skew). However, the domain generalization research space has segmented into learning-strategies and data augmentation approaches, the common thread being that neither branch is well-suited for prediction in many real-world enterprise settings. We propose a robust domain generalization framework capable of addressing training-serving-skew in unseen, noisy, structured and unstructured data environments routinely encountered in enterprise analytics environments, thereby allowing predictive models to be effective and valuable in settings constrained by data management capabilities. We evaluate our framework in three important contexts encompassing different data modalities: customer analytics (structured tabular), patient outcome forecasting (time series), and user modeling (text). When applied to unseen domains, our framework outperforms state-of-the-art methods, with performance gains often becoming even more pronounced on noisier data. Interestingly, the framework performs better than models trained within those domains. Importantly, cost-benefit analyses reveal that our prediction error costs are generally only 5-10% of those attained by the best comparison techniques, underscoring the downstream economic value of more generalizable predictive analytics in real-world enterprise settings. Our work has implications for IS research on the design and management of AI in organizations, and practical implications for managers looking to monetize predictive analytics in enterprise environments.

Keywords: Predictive Analytics, Domain Generalization, Machine Learning, Data Augmentation