LOCAL-OWA: neighbourhood-adaptive Ordered Weighted Averaging

Strengths

• •Combines two independent spatial modelling dimensions: range-sensitive local criterion weights (spatial heterogeneity) + global ORness (risk attitude)
• •Strictly generalises LOCAL-WLC (recovered when λ=1/n) and OWA (recovered when local ranges = global ranges)
• •Local weights are automatically derived from data: no extra subjective input beyond global w_k and λ
• •Score is in [0,1] by construction (normalised OWA)

Limitations

• •Neighbourhood definition choice affects local weights: critical and potentially subjective
• •Global λ applies the same risk attitude everywhere; if risk attitude must vary spatially, use SOWA instead
• •Denominator is alternative-specific (not neighbourhood-specific): subtle but important implementation detail
• •Requires at least 2 alternatives per neighbourhood; zero local range is a degenerate case

Assumptions to verify

• •Neighbourhood partition is meaningful (not arbitrary zone boundaries)
• •Each neighbourhood contains at least 2 alternatives
• •Range-sensitivity principle holds locally: criteria with wider neighbourhood spread should receive higher local weight
• •A single global risk attitude (ORness) is appropriate: if risk attitude must vary spatially, use SOWA
• •Global normalisation is appropriate (all alternatives in same measurement context)

When not to use

• •Non-spatial problem: no neighbourhood structure → use OWA
• •Only one alternative per neighbourhood → local range undefined
• •Risk attitude must vary spatially → use SOWA instead
• •All neighbourhoods have equal criterion spreads → standard OWA is sufficient

How to cite