Poor performance of KNeighborsClassifier for sklearn version >=0.24.2

Describe the bug

Hi. I used sklearn.neighbors.KNeighborsClassifier in my program but I noticed that its runtime increased when adopting higher versions of sklearn.

As shown in the following experiment results, this model performs best if we use sklearn <0.24.2.

My question is that why does such a library version perform better? Are there any issues in newer versions?

The detailed information is as follows:

Runtime	Memory	Version
451.2495478	2337.8830499649	1.0.1
441.4340003	2338.7937555313	0.24.2
26.25025550	414.1812286377	0.23.2
26.42619310	408.6705312729	0.22.1
26.42619310	408.6052923203	0.22
31.0837617	409.0471181870	0.21.3
29.0283316	408.5676021576	0.20.3
22.6652508	409.7873554230	0.19.2

My program only used the apis of KNeighborsClassifier provided by sklearn.

knn = KNeighborsClassifier(n_neighbors = 3)

Steps/Code to Reproduce

I create a minor program that used KNeighborsClassifier and you can test it on colab. The runtime is still higher when adopting the newer sklearn version on this example.

train.csv

Using the above train.csv and the following snippet:

# %%
# %%
import pandas as pd

train_data = pd.read_csv("train.csv")
total = train_data.isnull().sum().sort_values(ascending=False)
percent_1 = train_data.isnull().sum() / train_data.isnull().count() * 100
percent_2 = (round(percent_1, 1)).sort_values(ascending=False)
missing_data = pd.concat([total, percent_2], axis=1, keys=["Total", "%"])
cor = train_data.corr()
cor_target = abs(cor["target"])
relevant_features = cor_target[cor_target > 0.5]
s = train_data.dtypes == "object"
train_data_cat_var = list(s[s].index)
train_data.drop(
    ["nom_5", "nom_6", "nom_7", "nom_8", "nom_9", "ord_3", "ord_4", "ord_5"],
    axis=1,
    inplace=True,
)
train_data_cat_var = [
    ele
    for ele in train_data_cat_var
    if ele
    not in ["nom_5", "nom_6", "nom_7", "nom_8", "nom_9", "ord_3", "ord_4", "ord_5"]
]
final_train_data = pd.get_dummies(
    train_data, columns=train_data_cat_var, drop_first=True
)
features = final_train_data.drop(["target"], axis=1).columns
target = final_train_data["target"]

# %%
from sklearn.model_selection import train_test_split

X_train, X_test, y_train, y_test = train_test_split(
    final_train_data[features], target, test_size=0.2
)

# %%
from sklearn.neighbors import KNeighborsClassifier

# %%
%%time
knn = KNeighborsClassifier(n_neighbors=3, algorithm="brute")
knn.fit(X_train, y_train)
Y_pred_knn = knn.predict(X_test)

Expected Results

Similar runtime for the different sklearn versions or better performance for higher sklearn versions.

Actual Results

According to the above experiment results, when sklearn version is 0.24.2 or 1.0.1, the performance of the model is worse.

Versions

I test my program with sklearn 1.0.1, 0.24.2, 0.23.2, 0.22.1, 0.22, 0.21.3, 0.20.3 and 0.19.2.

[glemaitre]

Can you provide a minimal example that we can copy-paste. Right now the data are located to your drive.

Could you also try the last release of scikit-learn 1.1.1. We improve the scaling of the NN algorithm (even thought the regression shown seems really weird).

[DSOTM-pf]

Hello, Thanks for your quick reply. I test the snippet with different sklearn versions and different algorithms . The detailed information is as follows:

Python version	Sklearn version	Runtime(s)	Peak memory(MB)	Algo
3.8.13	1.1.1	89.47	2205.30	brute
3.8.13	1.0.2	90.31	2205.85	brute
3.7.10	1.0.2	86.89	2206.15	brute
3.8.13	1.1.1	3.86	233.28	ball_tree
3.8.13	1.0.2	4.35	232.80	ball_tree
3.7.10	1.0.2	3.56	233.05	ball_tree

I print the peak memory by tracemalloc.get_traced_memory() . The algorithm brute seems to cause the runtime and memory to increase.

Device info

OS : Ubuntu CPU : Intel® Core™ i9-9900K CPU GPU : TITAN V