Class-Based Attribute Weighting for Time Series Classification
Zolt ́ an · 2010
In this paper, we present two novel class-based weighting methods for the Euclidean nearest neighbor al- gorithm and compare them with global weighting methods considering empirical results on a widely accepted time se- ries classification benchmark dataset. Our methods provide higher accuracy than every global weighting in nearly half of the cases and they have better overall performance. We conclude that class-based weighting has great potential for improving time series classification accuracy and it might be extended to use with other distance functions than the Eu- clidean distance.