Music Album Review Sentiment Analysis

A machine learning system that classifies music album reviews into 6 sentiment categories using Bi-Directional LSTM, Gated Recurrent Unit and traditional ML models.

Python TensorFlow Keras Scikit-learn Pandas NLTK Matplotlib

Project Overview

This sentiment analysis system processes music album reviews and classifies them into 6 sentiment categories (0-5 stars) using advanced natural language processing techniques.

The project involved extensive data preprocessing, exploratory analysis, and implementation of both traditional machine learning models (Logistic Regression, SVM, Random Forest) and deep learning approaches (Bi-Directional LSTM). The system handles class imbalance through oversampling and achieves strong performance metrics.

We use BiDirectional LSTM over GRU even its accuracy is better than Bidirectional LSTM because Bidirectioanl LSTM can better perform on ambiguous sentences due to its dual traversal.

Key Innovation: The combination of traditional ML models with a Bi-Directional LSTM architecture provides robust sentiment classification, with the LSTM model particularly effective at capturing contextual relationships in review text.

Technical Highlights

Data Processing
  • 78,162 music album reviews
  • Handled class imbalance with oversampling
  • Text cleaning and tokenization
Model Architecture
  • Bi-Directional LSTM with embedding layer
  • Global max pooling
  • Dense layers with dropout
Evaluation
  • Accuracy, Precision, Recall metrics
  • Confusion matrix analysis
  • Comparison with traditional ML models
Technologies
  • TensorFlow/Keras for deep learning
  • Scikit-learn for traditional ML
  • Pandas/Numpy for data manipulation

Model Performance

Accuracy

86.4%

Bi-Directional LSTM
Precision

85.2%

Weighted average
Recall

86.4%

Weighted average
Model Comparison

The Bi-Directional LSTM outperformed traditional machine learning models:

  • Logistic Regression: 78.3% accuracy
  • Random Forest: 81.6% accuracy
  • SVM: 79.8% accuracy
Confusion Matrix
Figure 2: Confusion matrix showing classification performance across 6 sentiment classes
Project Links
View Source Code View Visualizations Download Report
📊 Project Details
  • Status: Completed
  • Dataset: 78,162 reviews
  • Classes: 6 sentiment levels
  • Framework: TensorFlow 2.x
Technology Stack
Core Libraries
TensorFlow Keras Scikit-learn Pandas Numpy
NLP
NLTK Tokenizer Word Embeddings
Visualization
Matplotlib Seaborn Plotly
Dataset Information
Original Distribution
  • 5-star: 29,520 (37.8%)
  • 4-star: 39,045 (49.9%)
  • 3-star: 4,430 (5.7%)
  • 2-star: 4,245 (5.4%)
  • 1-star: 525 (0.7%)
  • 0.5-star: 397 (0.5%)
After Oversampling

Balanced to 39,045 samples per class

Significant class imbalance required oversampling for model training

Data Visualizations

Rating Distribution

Original rating distribution showed extreme imbalance with most reviews being 4 or 5 stars. Oversampling was applied to create balanced classes for training.

Review Length Distribution

Review length analysis showed most reviews under 600 words. The 95th percentile was used to determine optimal padding length for the LSTM model.

Model Architecture

The Bi-Directional LSTM architecture with embedding layer, global max pooling, and dense layers with dropout achieved the best performance.

Technical Challenges & Solutions

The original dataset had extreme class imbalance with 49.9% 4-star and 37.8% 5-star reviews, while other classes were underrepresented.

Solution: Implemented oversampling to balance all classes to 39,045 samples each, ensuring the model wouldn't be biased toward majority classes.

Result: Model learned to recognize patterns in all sentiment classes equally.

Review lengths varied from a few words to thousands, making consistent model input challenging.

Solution: Analyzed length distribution and used the 95th percentile (612 words) as max length for padding/truncation.

Result: Balanced information retention with computational efficiency.

Needed to determine whether traditional ML or deep learning would perform better for this task.

Solution: Implemented and compared Logistic Regression, SVM, Random Forest, and Bi-Directional LSTM models.

Result: Bi-Directional LSTM achieved superior performance (86.4% accuracy vs 78-82% for traditional models).

Key Code Implementation

Bi-Directional LSTM Model Architecture
model = tf.keras.Sequential([
    tf.keras.layers.Embedding(input_dim=30000, 
                            output_dim=128, 
                            input_length=max_len),
    tf.keras.layers.Bidirectional(
        tf.keras.layers.LSTM(64, return_sequences=True)),
    tf.keras.layers.GlobalMaxPooling1D(),
    tf.keras.layers.Dense(64, activation='relu'),
    tf.keras.layers.Dropout(0.5),
    tf.keras.layers.Dense(6, activation='softmax')
])
Data Preprocessing
# Handle class imbalance through oversampling
dfs = [df[df['Rating_Class'] == i] for i in range(6)]
max_count = max(len(d) for d in dfs)
dfs_balanced = [resample(d, replace=True, 
                        n_samples=max_count, 
                        random_state=42) 
                for d in dfs]
df_balanced = pd.concat(dfs_balanced)
Text Tokenization
tokenizer = Tokenizer(num_words=30000, oov_token='')
tokenizer.fit_on_texts(df_balanced['Review'])
sequences = tokenizer.texts_to_sequences(df_balanced['Review'])
padded = pad_sequences(sequences, maxlen=max_len, 
                      padding='post', truncating='post')
Copyright © ynex. Designed with by Spruko All rights reserved