Shared nearest neighbors (SNN) measures neighborhood similarity by counting how many nearest neighbors two observations have in common. Points generated by the same mechanism should be close and share many neighbors. section 6.10
SNN supports neighborhood construction when ordinary distances become unreliable:
- Construct each point’s reference set from points with the most shared neighbors rather than the smallest Euclidean distances. section 6.10
- Use the resulting neighborhood within SOD to identify a subspace whose features explain the greatest variance for that reference set. section 6.10
- Rely on neighbor rank when high dimensionality or irrelevant features make specific distances less dependable. section 6.10
Correlated features in a 35-feature dataset
A synthetic dataset contains 100 records and 35 features; features 8 and 9 are correlated, while the others are irrelevant. One record is given an unusual combination of those correlated features. SOD uses a shared-neighbor reference set and identifies that record as the top outlier. section 6.10
