The default build algo for building the knn graph in UMAP is set to auto, which decides which algorithm to use to build the knn graph depending on the given dataset size (<= 50K uses brute force knn, >50K uses nn descent). Other options for build_algo are brute_force_knn and nn_descent, which uses that knn building algorithm without looking at the data size.
Related PR
However, to keep consistency for users who expect the same results as the previous release, we will use brute_force_knn to build the knn graph for the following case regardless of the data size.
random_state is explicitly set andbuild_algo = auto
TODO
Go through a deprecation cycle to let build_algo be auto (i.e. let the build algo be decided during runtime based on the data size) if given as auto regardless of whether random_state is explicitly given or not.
Description
The default build algo for building the knn graph in UMAP is set to
auto, which decides which algorithm to use to build the knn graph depending on the given dataset size (<= 50K uses brute force knn, >50K uses nn descent). Other options forbuild_algoarebrute_force_knnandnn_descent, which uses that knn building algorithm without looking at the data size. Related PRHowever, to keep consistency for users who expect the same results as the previous release, we will use
brute_force_knnto build the knn graph for the following case regardless of the data size.random_stateis explicitly set andbuild_algo = autoTODO
Go through a deprecation cycle to let
build_algobeauto(i.e. let the build algo be decided during runtime based on the data size) if given asautoregardless of whetherrandom_stateis explicitly given or not.