AdroitAnandAI / NaiveBayes

10-fold CV Naive Bayes Classifier on BoW and TF-IDF vectorized Reviews with Optimal Alpha

NaiveBayes

10-fold CV Naive Bayes Classifier on BoW and TF-IDF vectorized Reviews with Optimal Alpha

Amazon Fine Food Reviews Analysis

Data Source: https://www.kaggle.com/snap/amazon-fine-food-reviews

Objective:

To find Precision, Recall, F1 Score, Confusion Matrix, Accuracy of 10-fold cross validation Naive Bayes on vectorized input data, for BoW and TF-IDF featurizations. TPR, TNR, FPR and FNR is calculated for both. The most frequent words in positive and negative class also needs to eb found out

At a glance:

Random Sampling is done to reduce input data size and time based slicing to split into training and testing data. The metrics obtained by applying 10-fold cross validation with Naive Bayes using BoW and tf-idf featurizations are compared to find optimal alpha for smoothing. The accuracy for different alpha values are plotted. The performance metrics with both featurizations are computed. Most frequent words in both classes are also enumerated.

Laplace Smoothing:

Range of Alpha: taken values ranging from ** 10ˆ-6 to 10ˆ3** To try out different values of alpha, Geometric Progression is used as step function in the above alpha range.

Plots

Summary Table

Observations

1. F1 Score of "Count based BoW with Multinomial NB" is slightly higher than "Binary BoW with Bernoulli NB" method. This indicates the loss of information when the count vector is made binary.

2. Accuracy of TF-IDF method is 2% less than that we got from BoW method.

3. TNR and Recall of TF-IDF with Multinomial NB method is as high as 99%. This indicates the negative reviews are identified properly with only 1% false positive rate.

4. If we need a system with a prime requirement to correctly idenfity as many negative reviews as possible, then "Multinomial NB on tf-IDF" should be used (as TNR = 0.99)

5. TPR of TF-IDF with Multinomial NB method is low (26%). 74% of positive reviews are not identified correctly (FPR = 0.74)

6. TPR of "BoW with Multinomial NB" has the highest F1 Score, amongst all the three methods. Hence, Bag of Words featurization with multinomial Naive Bayes is the classifer of choice.

7. Naive Bayes algorithm is based on Bayes’ theorem with a strong (naive) conditional independence assumption between the features. Hence, working with w2v features which are completely dependent is not a good idea. Naive Bayes on W2V and tf-idf weighted W2v is not done for this reason.