About the project
What it is
ShiftGuard studies a practical problem: a model can look healthy when it is deployed even while the data it receives is changing. I built a repeatable experiment to test whether those changes can act as an early warning.
The project compares several distance measures across different datasets, classifiers, and types of shift. It also looks at false alarms and warning time, because a signal is only useful when it gives a clear and timely warning.
