~Note: should be merged into staging, but will compare against fix_891 while in draft.~
Fixes #1057.
Changelog
Removed get_group_matches
A bunch of methods have been refactored to comply with a faster df.groupby()-based approach instead
Removed max_pacf() method for finding maximum coefficients for partial autocorrelation. Was quite slow and had minimal impact on accuracy based on benchmarks
Removed STL usage in LightGBM and NeuralTS mixers plus associated helper methods
Benchmarks
Noticeable runtime speedups, between 2.5x and 16x depending on dataset size and amount of group combinations (if applicable). This is expected given the reduced computational complexity of all modified algorithms.
~Note: should be merged into staging, but will compare against
fix_891while in draft.~Fixes #1057.
Changelog
get_group_matchesdf.groupby()-based approach insteadmax_pacf()method for finding maximum coefficients for partial autocorrelation. Was quite slow and had minimal impact on accuracy based on benchmarksLightGBMandNeuralTSmixers plus associated helper methodsBenchmarks
Noticeable runtime speedups, between 2.5x and 16x depending on dataset size and amount of group combinations (if applicable). This is expected given the reduced computational complexity of all modified algorithms.