Drug review classifier

A multinomial Naive Bayes classifier for sentiment analysis over drug reviews.

Dataset source

Preprosessing

• The data is pre-divided into training and testing sets
• I ended up using three classes: negative (1-3), neutral (4-7), positive (8-10). The Numpy arrays containing the labels for training and is created from column "rating" and it has values -1, 0 and 0 respectively to the classes presented before.
• The Numpy arrays containing review data for training and testing are obtained from the column "review".

Feature-modelling

• Here I use as features the frequencies of each word in the dataset.
• Scikit-learn has a class CountVectorizer that converts reviews in form of text strings to feature vectors so I used it.

Classifier design

• I ended up using multinomial Naive Bayes classification as it is well known for being useful for text classification.

Evaluation metrics

• With the trained model I predicted the labels for the test reviews and with scikit-learn's accuracy score compared them with the testing labels.

Results

• The model's accuracy is currently around 70%