Indonesian Sentiment Analysis

A machine learning project that analyzes sentiment in Indonesian text using XGBoost and transformer-based models. The project provides tools to classify text as positive, negative, or neutral.

Overview

This project implements a sentiment analysis system that:

• Scrapes reviews from Google Play Store (specifically from the MyPertamina app)
• Processes and labels Indonesian text data
• Trains an XGBoost model using TF-IDF feature extraction
• Implements an easy-to-use inference interface

Features

• Text sentiment classification (positive, negative, neutral)
• Pre-processed data using TF-IDF vectorization
• Optimized XGBoost classifier with Bayesian hyperparameter tuning
• Simple API for sentiment prediction

Installation

1. Clone the repository:

git clone https://github.com/yourusername/indonesian-sentiment-analysis.git
cd indonesian-sentiment-analysis

2. Install the required dependencies:

pip install -r requirements.txt

3. Make sure you have the model files in the model/ directory:
   • xgb_sentiment_analysis_pipeline.joblib
   • label_encoder.joblib

Usage

Quick Start

from sentiment_analyzer.XGBoostSentimentAnalyzer import XGBoostSentimentAnalyzer

# Initialize the sentiment analyzer
sentiment_analyzer = XGBoostSentimentAnalyzer()

# Predict sentiment for a text
text = "aplikasi ini sangat bagus"
prediction = sentiment_analyzer(text)
print(f"Sentiment: {prediction}")

Using the Inference Notebook

The project includes a Jupyter notebook for easy inference:

1. Open notebook_inference.ipynb
2. Change the TEXT variable to your input text
3. Run the notebook cells to get a sentiment prediction

Dataset

The model was trained on reviews scraped from the MyPertamina app on Google Play Store:

• 30,000 reviews in Indonesian language
• Distribution: ~74% negative, ~24% positive, ~2% neutral
• Initial labeling performed using a pre-trained multilingual sentiment model

Model Details

Feature Extraction

• TF-IDF Vectorizer
• n-gram range: (1, 2)
• No IDF weighting

Classifier

• XGBoost with optimized hyperparameters:
  • Learning rate: 0.087
  • Max depth: 8
  • Number of estimators: 600
  • Subsample: 0.763
  • Colsample bytree: 0.594
  • Gamma: 2.46e-09

Performance

The model was evaluated using an 80/20 train/test split with stratification.

Project Structure

├── model/                   # Saved model files
├── notebook_inference.ipynb # Inference notebook
├── notebook_training_model.ipynb # Training notebook
├── sentiment_analyzer/      # Python package
│   └── XGBoostSentimentAnalyzer.py # Sentiment analyzer class
├── requirements.txt         # Project dependencies
└── README.md                # Project documentation

License

[Include your license information here]

Contributing

[Include contribution guidelines here]

Acknowledgments

• Initial sentiment labeling performed using lxyuan/distilbert-base-multilingual-cased-sentiments-student
• This project was built with scikit-learn, XGBoost, and other open-source libraries