Michele LaurelliBlog

Distribution shift

Distribution shift: A change in data distribution between training and deployment that can degrade performance. It includes changes in features, labels or conditional relationships. Segment monitoring and tests on recent data help distinguish a volume change from a generalization problem.

In italiano: Distribution ShiftLegge V · Overfitting, bias e collasso dei modelli

It includes changes in features, labels or conditional relationships. Segment monitoring and tests on recent data help distinguish a volume change from a generalization problem.

Examples

  • New acquisition devices
  • A changing customer population
  • Language data from another period