CIBC’s Techniques to Maintain AI Model Accuracy Amidst Dataset Shift

This article explores the critical challenge of dataset shift in artificial intelligence, where changes in data distribution between training and production environments degrade model performance. It highlights the expertise of CIBC’s Advanced Analytics and Artificial Intelligence team in addressing this issue through their participation in the Vector Institute’s Dataset Shift Project. This industry-academia collaboration was initiated in response to behavioral shifts caused by the global pandemic. The project aimed to equip participants with strategies for detecting and adapting to dataset shift, particularly covariate shift in cross-sectional data. Ali Pesaranghader, a Senior AI Research Scientist at CIBC, illustrates how such shifts impact various banking applications, including fraud detection and client experience. The team utilized house sales data from Iowa to demonstrate methods for identifying inconsistencies in variable contributions over time. By detecting these shifts, organizations can correct algorithms to restore prediction accuracy without needing costly re-training from scratch. The article underscores the importance of adaptive AI techniques in dynamic environments, showcasing CIBC’s proactive approach to maintaining robust AI systems amidst evolving market conditions and external disruptions.