Movie Review Sentiment Analysis

Question

Project Overview

This project focuses on building an end-to-end Machine Learning pipeline to analyze user sentiment from IMDB movie reviews. Unlike standard binary classifiers, this system incorporates Decision Logic to identify neutral feedback, providing a more nuanced recommendation: "Recommend," "Do Not Recommend," or "Maybe Watch."

---

How I Built It (Technical Workflow)

1. Advanced Text Preprocessing:
   Raw text is often noisy. I developed a cleaning pipeline using NLTK and Regex:

   • Noise Removal: Stripped HTML tags and non-alphabetic characters.
   • Normalization: Converted all text to lowercase for consistency.
   • Stopword Removal & Lemmatization: Removed common fillers (e.g., "the", "a") and reduced words to their dictionary root (e.g., "watched" becomes "watch") to reduce dimensionality.

2. Feature Engineering (TF-IDF):
   To translate text into a language machines understand, I used TF-IDF (Term Frequency-Inverse Document Frequency). I limited the vocabulary to the top 5,000 features to ensure the model focuses on the most impactful words while ignoring rare noise.

3. Model Implementation:
   I chose Linear SVC (Support Vector Classification). It is highly efficient for high-dimensional text data and provides a clear decision boundary between sentiments.

---

Challenges & Solutions

1. The Neutral Data Gap

• The Challenge: The IMDB dataset is strictly binary (Positive/Negative). This meant the model had never "seen" a neutral review, causing it to force-classify "average" movies as either good or bad.
• The Solution: Instead of using model.predict(), I utilized model.decision_function(). By calculating how close a review lies to the decision boundary (Thresholding), I was able to manually trigger a Neutral (0) classification for reviews that lacked strong positive or negative conviction.

2. Balancing Overfitting

• The Challenge: Initially, the model showed high training accuracy but lower performance on new data.
• The Solution: I leveraged L2 Regularization (inherent in Linear SVC). By monitoring the gap between training and testing scores, I ensured the model generalized well. Limiting the TF-IDF features also helped prevent the model from "memorizing" specific review noise.

---

How To Use This Project

1. Setup: Ensure scikit-learn, pandas, and nltk are installed in your Python environment.
2. Input: Pass a raw string (movie review) into the get_recommendation function.

3. Inference Example:

   review = "The visuals were stunning, but the pacing was a bit slow and mediocre."
   print(get_recommendation(review)) 
       
   # Result -> Score: 0.04 | Sentiment: 0 | Action: Maybe watch (Neutral)
   

Final Performance

• Test Accuracy: 88%
• Generalization: The model maintains a balanced F1-score across both classes, proving it is robust against bias and ready for real-world unseen data.

---