Sentiment Analysis on Movie Reviews

Overview

This project aims to predict the sentiment (positive or negative) of movie reviews using various machine learning models. Below is a summary of the different models and hyperparameters tested, along with their respective accuracies.

Models and Experiments

1. Random Forest Classifier

Model: RandomForestClassifier(n_estimators=100)
Method: get_preds(test_feat, train_feat, y_test, y_train, RandomForestClassifier(n_estimators=100))
Accuracy: 0.8358

Description: A Random Forest Classifier with 100 trees was used as the initial model. It provided a decent baseline accuracy.

2. Logistic Regression with L2 Penalty

Model: LogisticRegression(penalty='l2', max_iter=500, C=1, random_state=42)
Method: get_preds(test_feat, train_feat, y_test, y_train, LogisticRegression(penalty='l2', max_iter=500, C=1, random_state=42))
Accuracy: 0.8548

Description: A Logistic Regression model with L2 regularization showed better performance than the Random Forest Classifier.

3. Logistic Regression with Reduced Regularization Strength

Model: LogisticRegression(penalty='l2', max_iter=500, C=0.1, random_state=42)
Method: get_preds(test_feat, train_feat, y_test, y_train, LogisticRegression(penalty='l2', max_iter=500, C=0.1, random_state=42))
Accuracy: 0.875

Description: Reducing the regularization strength (C=0.1) improved the accuracy further.

4. Logistic Regression with Class Weights Balanced

Model: LogisticRegression(penalty='l2', max_iter=500, C=1, class_weight='balanced', random_state=42)
Method: get_preds(test_feat, train_feat, y_test, y_train, LogisticRegression(penalty='l2', max_iter=500, C=1, class_weight='balanced', random_state=42))
Accuracy: 0.855

Description: Applying class weights to balance the dataset did not significantly improve accuracy.

5. Logistic Regression with Liblinear Solver

Model: LogisticRegression(penalty='l2', max_iter=500, C=1, solver='liblinear', random_state=42)
Method: get_preds(test_feat, train_feat, y_test, y_train, LogisticRegression(penalty='l2', max_iter=500, C=1, solver='liblinear', random_state=42))
Accuracy: 0.8544

Description: Changing the solver to 'liblinear' had a negligible effect on accuracy.

6. Logistic Regression with Saga Solver (Unsuccessful Run)

Model: LogisticRegression(penalty='l2', max_iter=1000, C=0.5, solver='saga', class_weight='balanced', random_state=42)
Method: get_preds(test_feat, train_feat, y_test, y_train, LogisticRegression(penalty='l2', max_iter=1000, C=0.5, solver='saga', class_weight='balanced', random_state=42))
Status: Not able to run

Description: Attempted to use the 'saga' solver with a longer maximum iteration limit and balanced class weights. Unfortunately, this model configuration failed to execute properly.

Conclusion

The best model, based on accuracy, was the Logistic Regression with C=0.1, achieving an accuracy of 0.875. Future work could include further tuning, exploring different models, or addressing the issue with the 'saga' solver.