IMDB Movie Reviews Sentiment Analysis:

This project is a sentiment analysis model built to classify IMDB movie reviews as either positive or negative using the IMDB dataset. It uses various machine learning models and deep learning techniques to handle the text data.

Table of Contents:

• Overview
• Dataset
• Models
• Aspect-Based Sentiment Analysis
• Results
• Video Explanation

Project Video Demo:

Click To Watch the video demonstration

Overview:

This project performs binary classification on the IMDB dataset, where movie reviews are classified into positive (1) and negative (0) sentiments. It uses various machine learning algorithms and an LSTM-based deep learning model for comparison.

The steps involved:

1. Preprocess the data (tokenization, vectorization, and padding).
2. Build models using different algorithms.
3. Train the models.
4. Evaluate their performance.
5. Save the best model for future predictions.

Dataset:

The dataset used is the IMDB movie reviews dataset consisting of 50,000 reviews split equally between training and testing sets.

• Positive reviews: 25,000
• Negative reviews: 25,000

The dataset is split into training and validation sets for model training and evaluation.

Models:

We tested the following machine learning models on the dataset to determine their performance:

1. Naive Bayes

• Model: Multinomial Naive Bayes with TF-IDF vectorization.
• Accuracy: 85.55%

2. Logistic Regression

• Model: Logistic Regression with TF-IDF vectorization.
• Accuracy: 89.29%

3. Naive Bayes (Small Dataset)

• Model: Multinomial Naive Bayes on a subset of the dataset.
• Accuracy: 84.60%

4. Gradient Boosting

• Model: Gradient Boosting Classifier with TF-IDF vectorization.
• Accuracy: 80.35%

5. LSTM (Deep Learning Model)

• Model: An LSTM (Long Short-Term Memory) neural network to capture sequential dependencies in text data.
• Accuracy: 64.27%

Model Architecture (for LSTM):

1. Embedding Layer: Converts integer-encoded words into dense vectors.
2. LSTM Layer: Captures long-term dependencies in review sequences.
3. Dropout Layer: Helps prevent overfitting.
4. Dense Layer: Sigmoid-activated output layer for binary classification.

Aspect-Based Sentiment Analysis

In addition to traditional sentiment analysis, this project explores Aspect-Based Sentiment Analysis (ABSA), which focuses on identifying sentiments related to specific aspects of the movies. For example, a review might express a positive sentiment towards the acting but a negative sentiment towards the plot. This allows for more granular insights, such as:

• Improving Movie Attributes: By understanding which aspects of a movie are well-received and which are not, filmmakers and marketers can make informed decisions on what to emphasize in future projects.
• Targeted Recommendations: Users can receive recommendations based on specific attributes they care about (e.g., great cinematography or compelling storylines).
• Enhanced Customer Feedback: Businesses can better understand customer feedback and improve their products based on specific strengths and weaknesses highlighted in reviews.

Results

The LSTM model is a simple neural network architecture, achieving a lower accuracy compared to traditional machine learning models like Naive Bayes and Logistic Regression. Below is a summary of the model performances:

• Naive Bayes: 85.55%
• Logistic Regression: 89.29%
• Naive Bayes (Small Dataset): 84.60%
• Gradient Boosting: 80.35%
• LSTM: 64.27%

The best performing model in this case was Logistic Regression with an accuracy of 89.29%.

---

Review Quality Score (RQS)

In this analysis, we also incorporate a Review Quality Score (RQS), which measures the quality of reviews based on various factors such as length, sentiment strength, and engagement metrics (like the number of likes). The benefits of using RQS include:

• Quality Over Quantity: By focusing on reviews with higher RQS, the model can leverage more insightful data, leading to better predictions.
• Filtering Noise: Lower quality reviews can skew sentiment analysis, but using RQS helps filter out less informative reviews.
• Informed Model Training: Higher quality reviews contribute to more robust training data, potentially improving model accuracy.

---