Overview of Model Observability

The Model Observability Service provides deep insights into the performance and health of machine learning models in production. Moving beyond basic monitoring, the service calculates multifaceted metrics to provide a holistic view of model behavior over time.

Key Pillars

  • Performance Estimation: Using meta-learning and confidence scores, the system estimates performance drift even when ground truth labels / targets are missing. When ground truth becomes available, the service automatically compares actual results against previous estimates.

  • Data Drift Detection: The service monitors changes in input data distributions and inter-feature relationships that may lead to model degradation.

The service currently supports Regression (single-output), Binary Classification, or Multiclass Classification problems.

homepage

Getting Started

Start with the Tutorial.