Synapse
⌘+K
Synapse
PulseExploreClubsResearchersJournals
Instagram
HomeClubsExplore
May 17, 2026Open Access

Persistent Reliability Geometry: Empirical Identification of Stable Inference Neighbourhoods in Correlated Feature Spaces

View Full Paper
Ask AI
Bookmark
Share

Authors

DDEEPANSHU

Discussion

Loading...

Member takes

Overview

This work demonstrates how local neighbourhood reliability varies under data disturbance, implying metrics alone may not capture stability.

Key Points

  • The aim is to analyze local neighbourhood organization stability in machine learning models under realistic data disturbance.
  • Introduced the Neighbourhood Persistence Score (NPS) and Interpretability Instability Score (IIS) for empirical analysis.
  • Conducted experiments on synthetic correlated manifolds and real benchmark datasets.
  • Assessed the relationship between neighbourhood reliability and predictive performance in machine learning systems.
  • Demonstrated that neighbourhood reliability can degrade significantly before predictive performance does.
  • Highlighted that traditional accuracy metrics do not fully reflect explanation stability in deployed models.

Cite This Study

DEEPANSHU (2026) studied this question.

synapsesocial.com/papers/6a095c037880e6d24efe1e77https://doi.org/10.5281/zenodo.20211236
View Full Paper
Ask AI
Bookmark
Share