NNAWA: A Granular Nearest Neighbor Imputation Technique Based on Alpha-Weighted Average

Pasi Luukka, Jan Stoklasa, Mahinda Mailagaha Kumbure

Abstract: In this paper, we investigate the suitability of $\alpha {-}$ weighted averaging (AWA) operator in the $k$-nearest neighbor type of data imputation. AWA is a more recent type of aggregation operator, which is averaging in predefined sets and giving these sets importance weights. The final estimation is given by weighted average of these set averages. The idea behind it is that nearest neighbors belonging to the closest set are getting highest weights, neighbors belonging to the second closest set are given the second highest weight etc. In the paper, the AWA-based nearest neighbor data imputation method is introduced and tested with five real-world data. This way we get more granular approach to nearest neighbor imputation. Results show that with the new NNAWA we are able to produce results with smaller RMSEs.