Amazon Reviews Sentiment Analysis using NLP

Overview

This project implements a complete Natural Language Processing (NLP) pipeline to perform sentiment analysis on Amazon product reviews. The goal is to classify customer feedback as positive or negative using classical machine learning techniques.

The workflow covers:

• Exploratory Data Analysis (EDA)
• Text preprocessing and cleaning
• Feature extraction using Bag-of-Words
• Model training with multiple classifiers
• Performance evaluation using standard classification metrics

---

Dataset Description

The dataset consists of Amazon product reviews with structured fields, including:

• verified_reviews: Textual review content

• rating: Numerical rating (typically 1–5)

• feedback: Binary sentiment label

  • 1 → Positive
  • 0 → Negative

Key Observations:

• Strong class imbalance:

  • ~2800 positive reviews
  • ~250 negative reviews

• Reviews vary significantly in length

• Missing values exist in the verified_reviews column (handled during preprocessing)

---

Exploratory Data Analysis (EDA)

Several exploratory steps were performed to understand the dataset:

1. Data Inspection

• .info() used to check data types and missing values
• .describe() used for statistical summaries

2. Rating Distribution

• Count plots reveal skew toward higher ratings

3. Review Length Analysis

• A new feature length was created
• Histogram shows distribution of review lengths

4. Sentiment Distribution

• Significant imbalance between positive and negative classes

5. WordCloud Visualization

• Positive and negative reviews were visualized separately
• Helps identify commonly used words in each sentiment category

---

Text Preprocessing Pipeline

A custom preprocessing function (message_cleaning) was implemented with the following steps:

Steps:

1. Remove punctuation
2. Tokenization
3. Stopword removal using NLTK

Implementation Details:

• Uses Python’s string.punctuation
• Uses NLTK English stopwords
• Outputs a cleaned list of tokens for each review

Example transformation:

• Raw: "This product is AMAZING!!!"
• Cleaned: ["product", "amazing"]

---

Feature Engineering

Bag-of-Words Representation

The project uses CountVectorizer from Scikit-learn:

• Custom analyzer: message_cleaning
• Converts text into a sparse matrix of token counts
• Vocabulary extracted from the entire dataset

Result:

• High-dimensional sparse feature matrix
• Shape: (num_samples, num_unique_words)

---

Model Pipeline

Data Splitting

• Train-test split: 80% training / 20% testing

---

Models Implemented

1. Multinomial Naive Bayes

• Suitable for discrete word count features
• Assumes feature independence
• Fast and efficient baseline model

2. Logistic Regression

• Linear model for binary classification
• Learns weights for each feature
• Handles high-dimensional sparse data well

3. Gradient Boosting Classifier

• Ensemble method using decision trees
• Sequential learning to reduce errors
• More complex, captures non-linear relationships

---

Evaluation Metrics

Each model is evaluated using:

• Confusion Matrix
• Precision
• Recall
• F1-score
• Support

Visualization:

• Heatmaps used for confusion matrices via Seaborn

---

Results and Performance

1. Multinomial Naive Bayes

• Performs well on text classification tasks
• High accuracy on majority class (positive reviews)
• Slight bias due to class imbalance

Key Insight:

• Strong baseline model with fast training time

---

2. Logistic Regression

• Improved balance between precision and recall
• Better generalization compared to Naive Bayes

Key Insight:

• Handles feature interactions better than Naive Bayes

---

3. Gradient Boosting Classifier

• More flexible model capturing non-linearities
• Higher computational cost
• Risk of overfitting on imbalanced data

Key Insight:

• Performance depends heavily on hyperparameters and dataset balance

---

Comparative Analysis

Model	Strengths	Weaknesses
Naive Bayes	Fast, simple, strong baseline	Assumes independence
Logistic Regression	Balanced, interpretable	Linear decision boundary
Gradient Boosting	Powerful, flexible	Slower, sensitive to imbalance

---

Key Insights

• Class imbalance significantly affects performance, especially recall for negative reviews
• Bag-of-Words is effective but limited (no semantic understanding)
• Stopword removal improves signal-to-noise ratio
• Logistic Regression provides the best trade-off between simplicity and performance

---

Project Workflow Summary

1. Data Loading
2. Data Cleaning (handling nulls)
3. Exploratory Data Analysis
4. Feature Engineering (text → numerical)
5. Train-Test Split
6. Model Training (3 algorithms)
7. Evaluation using classification metrics
8. Visualization of results

---

Technologies Used

• Python
• Pandas & NumPy
• Matplotlib & Seaborn
• NLTK
• Scikit-learn
• WordCloud

---

Conclusion

This project demonstrates a complete end-to-end NLP classification pipeline for sentiment analysis using traditional machine learning methods. It highlights the importance of preprocessing, feature extraction, and model selection when working with textual data.

The implementation provides a strong foundation for more advanced NLP techniques such as:

• TF-IDF vectorization
• Word embeddings (Word2Vec, GloVe)
• Deep learning models (LSTM, Transformers)

---