I am currently working on an algorithm that aims to reduce dimensions and map data within the non-negative orthant. Subsequently, the mapped data is utilized as input for a classifier. The classifiers I have employed thus far are 'LogisticRegression (LR)' and 'MultinomialNB,' both of which are part of the sci-kit learn library. To optimize their performance, I have fine-tuned both algorithms using GridSearchCV.

According to the sci-kit learn documentation for MultinomialNB, this classifier is well-suited for classification tasks involving discrete features, such as word counts in text classification. Normally, the multinomial distribution expects integer feature counts, but fractional counts like tf-idf can also be effective in practice.

Considering this information, one might expect LR to outperform MultinomialNB. However, in practice, LR tends to achieve lower accuracy results and often suffers from overfitting (with a train accuracy of 100%) when the maximum test accuracy is attained.

I have made efforts to comprehend the reasons behind the potential superiority of MultinomialNB in this particular dataset. Among the resources I have explored, two posts, namely [1] and [2], have proven to be highly informative. Specifically, I have implemented the approach outlined in [2] and observed that my data adheres to a Gamma distribution. However, I have been struggling to establish a connection between these findings and the insights presented in [1] . I would greatly appreciate assistance from someone experienced in interpreting the classification results.



Your Answer

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