0.18.0 - 2023-06-26
DBStream will now only recluster on demand, rather than at every call to
predict_many method scikit-learn models wrapped with
raised an exception if the model had not been fitted on any data yet. Instead, default predictions will be produced, which is consistent with the rest of River.
didn't check whether features were ordered in the same way at each method call. They now store the list of feature names at the first function call, and align subsequent inputs in the same order.
to power-up approximate nearest neighbor search. SWINN uses graphs to speed up nearest neighbor search in large sliding windows of data.
Standardize and create base classes for generic nearest neighbor search utilities.
The user can now select the nearest neighbor search engine to use in
min_branch_fraction parameter to avoid splits where most of the data goes to a single branch. Affects
max_share_to_split parameter to Hoeffding Tree classifiers. This parameters avoids splitting when the majority
class has most of the data.