SentimentAnalysis

Logistic Regression: A sentiment analysis case study

Key Concepts

• Build and employ a logistic regression classifier using scikit-learn
• Clean and pre-process data
• Perform feature extraction with nltk
• Tune model hyperparameters and evaluate model accuracy

Dataset

IMDB movie reviews dataset http://ai.stanford.edu/~amaas/data/sentiment Contains 25000 positive and 25000 negative reviews Contains at most reviews per movie At least 7 stars out of 10 : positive (label = 1) At most 4 stars out of 10 : negative (label = 0) 50/50 train/test split Evaluation accuracy

Features: bag of 1-grams with TF-IDF values:

Extremely sparse feature matrix - close to 97% are zeros

Model: Logistic regression

p(y=1|x)=σ(w.Tx)

Linear classification model Can handle sparse data Fast to train Weights can be interpreted