In plain words
The classic trap in outlier detection: a global density threshold fails the moment your data has both dense cities and sparse villages — everything in the sparse region looks anomalous. The fix is to judge each point relative to its own neighbourhood.
RKDOS computes a relative kernel-density outlier score using weighted kernel density estimation with an adaptive Gaussian kernel width, evaluated over an extended neighbourhood that includes both a point's k-nearest neighbours and its reverse nearest neighbours — the points that consider it a neighbour.
On synthetic and real datasets, RKDOS reported better detection than popular existing density-based approaches, and the work won the Best Paper Award at SKIMA 2018, later extended in IETE Technical Review.
How it works
- Estimate local densityWeighted kernel density estimation smooths each point's neighbourhood, with the Gaussian kernel width adapting to local sparsity.
- Look both waysThe neighbourhood includes k-nearest neighbours and reverse nearest neighbours, stabilising the estimate at cluster borders.
- Score relativelyA point's density is compared to its neighbours' densities — a low ratio, not a low absolute value, marks the outlier.
- Rank and flagPoints with the lowest relative scores surface as outliers, robust to datasets of mixed density.
Why it matters
Outlier detection underpins fraud screens, sensor-fault monitors and data-quality pipelines. A score that stays honest when density varies — the normal condition of real data — is a genuinely reusable building block, which is why this algorithmic work still gathers citations.