M nearest points to centroid in K-Means clustering

Question

I have implemented a function to find the nearest data point to each centroid calculated after running the K-Means clustering algorithm. I wanted to know if there's a sklearn function that allows me to find the M nearest points to each of the centroids.

Miculek

Answer 1

After running K-means, we can use sklearn.neighbors.NearestNeighbors to fit our dataset. We can then query the Nearest Neighbor model with our K-means centroids to retrieve the neighbors. Like this:

# Copyright 2024 Google LLC.
# SPDX-License-Identifier: Apache-2.0
from sklearn.cluster import KMeans
from sklearn.neighbors import NearestNeighbors

# random dense embeddings for 100 points with 10 dimensions.
dataset = np.random.rand(100,10)

# fit K-means with 3 clusters on our dataset.
kme = KMeans(n_clusters=3)
kme.fit(dataset)

# we should have 3 vectors for 3 centroids.
print(kme.cluster_centers_.shape) # (3, 10)

# initialize NearestNeighbor with 5 neighbors and fit our dataset.
knn = NearestNeighbors(n_neighbors=5, metric='cosine')
knn.fit(dataset)

# Use the model to query the centroids' neighbors.
distances, indices = knn.kneighbors(kme.cluster_centers_)

for centroid, distance_from_centroid, index in zip(kme.cluster_centers_, distances, indices):
  print(centroid, distance_from_centroid, index)

The last loop will output 3 lines. Each one resembles the vector of a centroid along with 5 distances and indices of its closest neighbors.

Eggett

Answer 2

Yes you want to check out the following tutorial from sklearn : http://scikit-learn.org/stable/modules/neighbors.html

The classsklearn.neighbors.NearestNeighbors finds it for you: http://scikit-learn.org/stable/modules/generated/sklearn.neighbors.NearestNeighbors.html#sklearn.neighbors.NearestNeighbors

Danner

eventhough this answer does not exactly answer the question, the point he makes is valid. You can use sklearn.neighbors.NearestNeighbors and use kneighbors function to find the nearest neighbours of the cluster centroids. – Clinkscales