I am having some difficulties in improving results from running a Naive Bayes algorithm. My dataset consists of 39 columns (some categorical, some numerical). However I only considered the main variable, i.e. Text, which contains all the spam and ham messages.

Since it is a spam filtering, I think that this field can be good. So I used countvectorizer and fit transform using them after removing stopwords.

I am getting a 60% of accuracy which is very very low! What do you think may cause this low result? Is there anything that I can do to improve it?

These are the columns out of 39 that I am considering:

Index(['Date', 'Username', 'Subject', 'Target',  'Country', 'Website','Text', 'Capital', 'Punctuation'],

Date is in date format (e.g. 2018-02-06) Username is a string (e.g. Math) Subject is a string (e.g. I need your help) Target is a binary variable (1 -spam or 0-not spam) Country is a string (e.g. US) Website is a string (e.g. www.viagra.com) Text is the corpus of the email and it is a string (e.g. I need your HELP!!) Capital is a string (e.g. HELP) Punctuation is string (!!)

What I have done is the following:

  • removing stopwords in Text:

    def clean_text(text):

      lim_pun = [char for char in string.punctuation if char in "&#^_"]
      nopunc = [char for char in text if char not in lim_pun]
      nopunc = ''.join(nopunc)
      other_stop=['•','...in','...the','...you\'ve','–','—','-','⋆','...','C.','c','|','...The','...The','...When','...A','C','+','1','2','3','4','5','6','7','8','9','10', '2016',  'speak','also', 'seen','[5].',  'using', 'get',  'instead',  "that's",  '......','may', 'e', '...it', 'puts', '...over', '[✯]','happens', "they're",'hwo',  '...a', 'called',  '50s','c;', '20',  'per', 'however,','it,', 'yet', 'one', 'bs,', 'ms,', 'sr.',  '...taking',  'may', '...of', 'course,', 'get', 'likely', 'no,']
      clean_words = [word for word in nopunc.split() if word.lower() not in ext_stopwords]
      return clean_words

Then applying these changes to my dataset:

from sklearn.feature_extraction.text import CountVectorizer
import string
from nltk.corpus import stopwords

df=df.dropna(subset=['Subject', 'Text']) 
mex = CountVectorizer(analyzer=clean_text).fit_transform(df['Corpus'].str.lower())

and split my dataset into train and test:

X_train, X_test, y_train, y_test = train_test_split(mex, df['Target'], test_size = 0.80, random_state = 0)

df includes 1110 emails with 322 spam emails.

Then I consider my classifier:

# Multinomial Naive Bayes 

from sklearn.naive_bayes import MultinomialNB

classifier = MultinomialNB()
classifier.fit(X_train, y_train)



# Train data set

    from sklearn.metrics import classification_report,confusion_matrix, accuracy_score
    from sklearn.metrics import accuracy_score
    pred = classifier.predict(X_train)
    print(classification_report(y_train ,pred ))
    print('Confusion Matrix: \n',confusion_matrix(y_train,pred))
    print("MNB Accuracy Score -> ",accuracy_score(y_train, pred)*100)
    print('Predicted value: ',classifier.predict(X_test))
    print('Actual value: ',y_test.values)

and evaluate the model on the test set:

from sklearn.metrics import classification_report,confusion_matrix, accuracy_score

pred = classifier.predict(X_test)

print(classification_report(y_test ,pred ))
print('Confusion Matrix: \n', confusion_matrix(y_test,pred))
print("MNB Accuracy Score -> ",accuracy_score(y_test, pred)*100)

getting approx 60%, which is not good at all. Output:

  precision    recall  f1-score   support

         0.0       0.77      0.34      0.47       192
         1.0       0.53      0.88      0.66       164

    accuracy                           0.59       356
   macro avg       0.65      0.61      0.57       356
weighted avg       0.66      0.59      0.56       356

Confusion Matrix: 
 [[ 66 126]
 [ 20 144]]

I do not know if the problem are the stopwords or the fact that I am considering only Text or Corpus as column (it would be also good to consider Capital letters and punctuation as variables in the model).

  • $\begingroup$ Please give more detail about your process: how many instances? Size of documents? What are the 39 features? How many words features? It might also be useful to look at precision/recall instead of only accuracy. $\endgroup$
    – Erwan
    Commented Sep 7, 2020 at 11:10
  • $\begingroup$ Hi Erwan, I updated the question. I think I answered to all your questions. Please let me know if you need more info. Thank you so much $\endgroup$
    – Math
    Commented Sep 7, 2020 at 11:31

1 Answer 1


Your model certainly overfits. It's likely that the main issue is the inclusion in the features of words which appear very rarely (especially those which appear only once in the corpus):

  • Words which appear only once don't help classification at all, if only because they can never be matched again. More generally, words which appear rarely are more likely to appear by chance, so using them as features causes overfitting.
  • Naive Bayes is very sensitive to overfitting since it considers all the features independently of each other.
  • It's also quite likely that the final number of features (words) is too high with respect to the number of instances. A low ratio instances/words causes overfitting.

The solution is to filter out words which occur less than $N$ times in the data. You should try with several values of $N$, starting with $N=2$.

Another issue: in your current process the data is preprocessed before splitting between training and test set, this can cause data leakage. Note that filtering out words of low frequency should be done using the training data only, and then just selecting the same words on the test set (ignoring any other word).


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.