A gradual decline in model accuracy after a period of strong performance is a hallmark of data drift, where the statistical properties of incoming input data shift over time relative to the data the model was originally trained on. Since the model's learned patterns are based on historical data, changes in the real-world input distribution (e.g., shifting user behavior, seasonal trends, or evolving data sources) cause the model's predictions to become less aligned with current reality. Azure Machine Learning provides data drift monitoring capabilities specifically to detect this kind of degradation. Insufficient compute would affect training/inference speed rather than accuracy, and incorrect training would typically produce poor results from the outset rather than a delayed decline.
Community Discussion