<ul><li>A significant challenge in maintaining real-world machine learning models is determining when to retrain or update the model due to continuous and unpredictable data evolution.</li><li>This decision is complex due to limited information availability, unknown nature of distribution shifts, and the need to specify a cost ratio between retraining and poor performance.</li><li>Existing methods do not provide a comprehensive solution to this retraining problem, as they fail to account for cost trade-offs, data scarcity, and key practical considerations.</li><li>A proposed uncertainty-based method continually forecasts model performance evolution using a bounded metric, outperforming existing baselines in experiments across 7 datasets for classification tasks.</li></ul>

When to retrain a machine learning model

Discover more