This project implements a neural network to perform sentiment analysis on book reviews from Amazon. The goal is to classify reviews as positive or negative based on their content.
The dataset contains book reviews along with labels indicating whether the review is positive or negative. It is loaded from a CSV file named bookReviews.csv.
To run this project, you need to have Python installed along with the following libraries:
- pandas
- numpy
- matplotlib
- seaborn
- scikit-learn
- tensorflow
- keras
The neural network consists of the following layers:
- Input Layer: Configured with input shape corresponding to the TF-IDF vectorized features.
- Hidden Layers:
- Four hidden layers with 128, 64, 32, and 16 units respectively, each using the ReLU activation function.
- Batch Normalization and Dropout layers added for regularization and improved generalization.
- Output Layer: A single unit with a sigmoid activation function for binary classification.
Conducted several experiments to optimize the model's performance:
- Tested learning rates: 0.01, 0.001, 0.0001
- Best result: Learning rate of 0.001
Experiment 2: Adding More Hidden Layers and Adjusting Units
- Configurations: (64, 32), (64, 32, 16), (128, 64, 32, 16)
- Best result: Configuration with four hidden layers (128, 64, 32, 16 units)
- Configurations:
min_df=1, ngram_range=(1, 1),min_df=2, ngram_range=(1, 1),min_df=1, ngram_range=(1, 2) - Best result:
min_df=1, ngram_range=(1, 2)
- Test Accuracy: 0.7932
- The final model configuration provided the highest test accuracy and best generalization.
The optimal configuration for the sentiment analysis model is:
- Learning Rate: 0.001
- Layer Configuration: Four hidden layers with 128, 64, 32, and 16 units respectively.
- Vectorizer Configuration:
min_df=1, ngram_range=(1, 2)
This setup achieved the highest test accuracy and provided a good balance between model complexity and performance.
This project is licensed under the MIT License.