I am trying to implement k-means clustering on 60-70 features and I came across a post for feature selection technique on quora by Julian Ramos, but I fail to understand few steps mentioned. I am also wondering if its the right method to select the best features for clustering?

These are the steps mentioned in the post :

Sf={∅} #Set of features selected, initially empty

  1. Perform k-means on each of the features individually for some k.
  2. For each cluster measure some clustering performance metric like the Dunn's index or silhouette.
  3. Take the feature which gives you the best performance and add it to Sf
  4. Perform k-means on Sf and each of the remaining features individually
  5. Take the feature which gives you the best performance and add it to Sf
  6. If you have reached the desired number of features stop, else go back to 4

Also, how do we implement the same in python. I wish to write function for the same that selects best k and implement all the other steps.

  • 1
    $\begingroup$ The approach you describe seems fine, it's basically greedy bottom up selection. Can you please elaborate which exact steps are you having trouble with? $\endgroup$
    – Vlad_Z
    Commented Jan 25, 2020 at 17:54
  • 4
    $\begingroup$ Be careful: feature selection with unsupervised methods is risky because the algorithm will favor features which are easy to cluster and discard harder features even though they might be meaningful for the task. $\endgroup$
    – Erwan
    Commented Jan 26, 2020 at 0:24
  • $\begingroup$ @Vlad_Z I am having trouble understanding the step 4. $\endgroup$ Commented Jan 26, 2020 at 16:21

4 Answers 4


Often people confuse unsupervised feature selection (UFS) and dimensionality reduction (DR) algorithms as the same. For instance, a famous DR algorithm is Principal Component Analysis (PCA) which is often confused as a UFS method! Researchers have suggested that PCA is a feature extraction algorithm and not feature selection because it transforms the original feature set into a subset of interrelated transformed features, which are difficult to emulate (Abdi & Williams, 2010).

A UFS approach present in literature is Principal Feature Analysis PFA. The way it works is given as; Steps:

  • Compute the sample covariance matrix or correlation matrix,
  • Compute the Principal components and eigenvalues of the Covariance or Correlation matrix A.
  • Choose the subspace dimension n, we get new matrix A_n, the vectors Vi are the rows of A_n.
  • Cluster the vectors |Vi|, using K-Means
  • For each cluster, find the corresponding vector Vi which is closest to the mean of the cluster.

A possible python implementation of PFA is given below;

from sklearn.decomposition import PCA
from sklearn.cluster import KMeans
from collections import defaultdict
from sklearn.metrics.pairwise import euclidean_distances
from sklearn.preprocessing import StandardScaler
import pandas as pd
# create some dummy data
df = pd.DataFrame({'num_legs': [2, 4, 8, 0],
                   'num_wings': [2, 0, 0, 0],
                   'num_specimen_seen': [10, 2, 1, 8]},
                  index=['falcon', 'dog', 'spider', 'fish'])
class PFA(object):
    def __init__(self, n_features, q=None):
        self.q = q
        self.n_features = n_features
    def fit(self, X):
        if not self.q:
            self.q = X.shape[1]
        sc = StandardScaler()
        X = sc.fit_transform(X)
        pca = PCA(n_components=self.q).fit(X) # calculation Covmatrix is embeded in PCA
        A_q = pca.components_.T
        kmeans = KMeans(n_clusters=self.n_features).fit(A_q)
        clusters = kmeans.predict(A_q)
        cluster_centers = kmeans.cluster_centers_
        dists = defaultdict(list)
        for i, c in enumerate(clusters):
            dist = euclidean_distances([A_q[i, :]], [cluster_centers[c, :]])[0][0]
            dists[c].append((i, dist))
        self.indices_ = [sorted(f, key=lambda x: x[1])[0][0] for f in dists.values()]
        self.features_ = X[:, self.indices_]
# Usage
pfa = PFA(n_features=3)
# To get the transformed matrix
x = pfa.features_
# To get the column indices of the kept features
column_indices = pfa.indices_


        num_legs  num_wings  num_specimen_seen
falcon         2          2                 10
dog            4          0                  2
spider         8          0                  1
fish           0          0                  8
[[-0.50709255  1.73205081  1.23942334]
 [ 0.16903085 -0.57735027 -0.84802649]
 [ 1.52127766 -0.57735027 -1.10895772]
 [-1.18321596 -0.57735027  0.71756088]]


Abdi, H., & Williams, L. J. (2010). Principal component analysis. Wiley interdisciplinary reviews: computational statistics, 2(4), 433-459.


You might be interested in this paper and python implementation of various other feature selection for clustering tools and papers:



An excerpt sumarizing the approach:

We address the problem of selecting a subset of important features for clus tering for the whole data and not just for clusters unlike in [1,2] This helps in knowing the important features before doing clustering and the clustering task becomes more ecient and focused as only the important features can be used Finding the important original features for the whole data helps in under standing the data better unlike principal components Data storage collection and processing tasks become more efficient and noise is reduced as the data is pruned

Our approach is a 2-step method we first rank and then select a subset of important features. Ranking of features is done according to their importance on clustering An entropy based ranking measure is introduced We then select a subset of features using a criterion function for clustering that is invariant with respect to different numbers of features A novel scalable method based on random sampling is introduced for large data commonly found in data mining applications

Additionally, this paper gives an overview of different methods available:



according to your citated discription it looks like Hierarchical Clustering - you can see for it in scikit-learn lib python


I don't recommend using PCA before clustering because it makes labelling clusters very difficult. PCA groups different features together and while this does reduce number of features, the resulting features are very different from the original and will likely make your cluster center naming very difficult.

I think using simple statistical based filtering would suffice:

  1. check variance-covariance matrix to see if there are features that highly correlated and decide which to keep based on business knowledge

  2. check the variance of features. Features with very low variance will likely not contribute meaningfully to clustering.

  3. if there are categorical features that have many components, consider condensing the sparse components into "other" to avoid creating too many features when performing encoding.

For auto selecting k in K-Means, there are actually quite a few ways you can auto select best k. One way is by silhouette score (select highest) or you can create an adjusted inertia formula that adds a penalty term for clusters.

link to some post with their different approaches: https://medium.com/analytics-vidhya/how-to-determine-the-optimal-k-for-k-means-708505d204eb


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.