'''I used the k-means algorithm to clustering set of documents which are textual data only.

The document has 2lack records.

Surprisingly the result for the clustering is

90% of records is storing in 1 cluster, remaining records are going to another clusters.

It's is not a problem if it chooses correct cluster while predicting the data. I thought uneven distribution of clusters might be the issue behind my problem. If thats not the problem, Suggest me how k-means will predict the correct cluster.

Why this is happening ?

Parameters : number of clusters = 5 (used elbow method to know this), random state = 0 or 42 (both used but no use)

#Here is my code showing how i created vectors

from sklearn.feature_extraction.text import TfidfVectorizer
import _pickle as cPickle

def build_tfidf_vect(series,save_model = True ,) :
    vectorizer = TfidfVectorizer(stop_words="english")
    vectors = vectorizer.fit_transform(series)
    print("Shape of tfidf matrix: {}".format(vectors.shape))
    if save_model:
        data_struct = {'vectors': vectors, 'vectorizer': vectorizer}
        with open('data_2l.bin', 'wb') as f:
            cPickle.dump(data_struct, f)
    return vectorizer, vectors

import pandas as pd

feed = pd.read_csv('2l_data.csv', encoding='latin')


This is how i created clusters.

from sklearn.cluster import MiniBatchKMeans
import scipy

def load_tf_idfvectors():
    import time
    import _pickle as cPickle
    with open(r'C:\Users\data_2l.bin', 'rb') as f:
        data_struct = cPickle.load(f)
        vectors,vectorizer = data_struct['vectors'], data_struct['vectorizer']
    return vectorizer,vectors

def dump(cluster_0,cluster_1,cluster_2,cluster_3,cluster_4):
    if save_model:
        data_struct = {'cluster0': cluster_0, 'cluster1': cluster_1,'cluster2': cluster_2, 'cluster3': cluster_3,'cluster4': cluster_4}
        with open(r'C:\Users\totalclusters_2l.bin', 'wb') as f:
            cPickle.dump(data_struct, f)

import pandas as pd
import pickle
data=pd.read_csv(r"C:\Users\New_vectors_data_2l.csv", encoding="latin")

kmeans = MiniBatchKMeans(n_clusters=5, init= 'k-means++',random_state=0).fit(tfidf)
labels = kmeans.fit_predict(tfidf)
result = pd.concat([data['Column_name'],X_n,labels_n])
pickle.dump(kmeans, open(filename, 'wb'))
results_0=result.loc[result['cl_n'] ==0 , ['tf-idf','cl_n']]
cluster_0_tf = scipy.sparse.vstack(results_0['tf-idf'].tolist())
open_file = open(r'C:\Users\cluster0_2l', "wb")
pickle.dump(cluster0_in, open_file)

  • $\begingroup$ This can happen if you have outliers which form their own clusters. $\endgroup$ Dec 16, 2022 at 11:38
  • $\begingroup$ ok @user2974951 but while predicting the data, kmeans is not picking(predicting) correct cluster which has the exact data. why it is happening ? because of uneven cluster distribution or outlier clusters. Suggest me some solution... $\endgroup$
    – sai_0033
    Dec 16, 2022 at 11:52

1 Answer 1


From what I know, MiniBatchKMeans has a changing 'predict' function by design.
In every batch the centeriods are calculated based on data until this point.
Meaning, that the same input can get different predicted label(cluster), depending on the batch it was in.
See this answer.

A tradeoff between supporting an ongoing streamed data(letting the model dynamically learn) and having constant results per input(since the model has only one learning context -- one 'fit' operation)


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.