[BUG] Pipeline of MinMaxScaler and KMeans with DataFrame input produces incorrect results

[tfeher]

Describe the bug When using a Pipeline([ ('minmax', MinMaxScaler()), ('kmeans', Kmeans())]), with a cudf DataFrame input, the results will be incorrect if the dataframe index is not sorted.

Steps/Code to reproduce bug

#!/usr/bin/env python
# coding: utf-8

# In[2]:


import cudf
import numpy as np
import cupy as cp
import matplotlib.pyplot as plt
from cuml.cluster import KMeans
from cuml.pipeline import Pipeline
from cuml.preprocessing import MinMaxScaler
from cuml.datasets import make_blobs


# In[4]:


# Create dataset of random blobs
n_rows = 1000
n_cols = 2
dataset, _ = make_blobs(n_samples=n_rows, n_features=n_cols, centers=2, cluster_std=1.0, center_box=(-10.0, 10.0), random_state=137, dtype='float32')

# Create dataframe with random index
idx = cp.random.permutation(n_rows)
df = cudf.DataFrame(data={"I": idx, "A": dataset[:,0], "B": dataset[:,1]})
df = df.set_index("I")
print(df)


# In[5]:


# Clustering the cupy the dataset works as expected
mms = MinMaxScaler()
kmeans = KMeans(n_clusters=2)
pipeline = Pipeline([
       ('minmax', mms),
       ('kmeans', kmeans)
    ])
pipeline.fit(dataset)
labels = pipeline.predict(dataset)

plt.scatter(cp.asnumpy(dataset[:,0]), cp.asnumpy(dataset[:,1]), c=cp.asnumpy(labels))


# In[6]:


# Using the dataframe as input does not work
mms = MinMaxScaler()
kmeans = KMeans(n_clusters=2)
pipeline = Pipeline([
       ('minmax', mms),
       ('kmeans', kmeans)
    ])

pipeline.fit(df[['A', 'B']])

prediction = pipeline.predict(df[['A', 'B']])
df['C'] = prediction

plt.scatter(df["A"].to_numpy(), df["B"].to_numpy(), c=df["C"].to_numpy())


# In[7]:


# If the dataframe is sorted then it works
df = df.sort_index()
mms = MinMaxScaler()
kmeans = KMeans(n_clusters=2)
pipeline = Pipeline([
       ('minmax', mms),
       ('kmeans', kmeans)
    ])

pipeline.fit(df[['A', 'B']])

prediction = pipeline.predict(df[['A', 'B']])
df['C'] = prediction
# tmp = df.reset_index()[['I', 'C']]

plt.scatter(df["A"].to_numpy(), df["B"].to_numpy(), c=df["C"].to_numpy())

Expected behavior I expect to run the pipeline and get correct clustering regardless whether the dataframe index is sorted or not.