Skip to content

Provenance and Uncertainty

Anubha Parashar edited this page Aug 10, 2026 · 1 revision

Provenance and Uncertainty

IncidentGraph treats provenance, contradiction and missingness as first-class parts of the output rather than optional explanations added after inference.

Provenance faithfulness

Accepted event edges are required to contain supporting evidence references. This keeps a path from a graph claim back to the camera/source observations that contributed to it.

Non-negative missingness

If a sensor is unavailable over an interval, IncidentGraph records a missing-evidence object and leaves affected transitions unresolved.

Not observed ⇒ unknown, not did not occur.

This prevents a common logical error in sparse or asynchronous camera networks.

Contradiction preservation

Competing claims are grouped. The selected claim can be identified for downstream use, but losing subjects, scores and proposals remain available for audit.

Calibration

The activity-recognition study reports expected calibration error (ECE) and multiclass Brier score. Temperature scaling is used as a post-hoc classifier calibration baseline.

Graph-level confidence

The controlled benchmark also evaluates association-confidence calibration. Its aggregate diagnostic graph score includes entity/event quality, temporal order, contradiction handling, missing-evidence handling, provenance coverage and calibration.

Warning

The aggregate graph score is an engineering diagnostic average, not a learned or validated decision utility for deployment.

Clone this wiki locally