🌟 Analyzing Customer Perceptions of Products and Services at Leading Indonesian Coffee Shops Using CNN-BiLSTM and Zero-Shot Learning

📄 Project Overview

This project delivers a comprehensive sentiment analysis system targeting customer opinions about four well-known coffee shop brands in Indonesia: Fore Coffee, Kopi Kenangan, Point Coffee, and Tomoro Coffee. Leveraging a deep learning model that combines Convolutional Neural Networks (CNN) with Bidirectional Long Short-Term Memory (BiLSTM) networks, the system classifies tweets into three sentiment categories: positive, negative, and neutral.

🧾 Abstract

The rapid growth of the coffee shop industry in Indonesia has made understanding customer perceptions crucial for maintaining a competitive edge. Social media platforms like Twitter offer rich and dynamic sources of public opinion; however, the unstructured nature of these texts necessitates advanced analytical methods. This study conducts a comparative sentiment analysis of customer perceptions toward two leading coffee shop brands—Fore Coffee and Kopi Kenangan—using Twitter data collected throughout 2024. The methodology integrates a hybrid Convolutional Neural Network–Bidirectional Long Short-Term Memory (CNN-BiLSTM) model for sentiment classification and a Zero-Shot Learning approach with a large language model (LLM) for aspect-based sentiment extraction. A total of 2,277 tweets related to Fore Coffee and 1,593 tweets related to Kopi Kenangan were analyzed. The results indicate a predominantly positive perception for both brands, with Fore Coffee receiving 71.6% positive sentiment and Kopi Kenangan 72.4%. Product quality emerged as the key driver of positive sentiment for both. Nonetheless, strategic differences were identified. Fore Coffee is perceived to offer consistently premium-quality products but is challenged by a perception of high prices. In contrast, Kopi Kenangan is appreciated for affordability and product customization, though concerns arise regarding product consistency and service accuracy. Based on these findings, tailored strategic recommendations are proposed: Fore Coffee should reinforce its value proposition to justify its premium pricing, while Kopi Kenangan is advised to focus on operational standardization and improving service accuracy to sustain its rapid expansion.

🤖 Key Features

• Twitter data collection based on brand-specific keywords
• Text preprocessing: normalization, stemming, stopword removal
• Word embedding using pre-trained GloVe vectors
• Hybrid CNN-BiLSTM deep learning architecture for sentiment classification
• Evaluation using 10-fold cross-validation for robustness
• Visualization of performance metrics and sentiment word clouds

🔧 Technologies Used

• Python: NumPy, Pandas, Scikit-learn
• TensorFlow / Keras for deep learning modeling
• NLTK, Sastrawi, Spacy, TextBlob for NLP preprocessing
• Matplotlib, Seaborn, WordCloud for data visualization
• Zero-Shot Learning for aspect extraction

📁 Project Structure

.
├── data/                  # Datasets for the project
│   ├── raw/               # Raw, immutable data
│   └── output/            # Processed data
├── models/                # Trained and serialized models
├── notebooks/             # Jupyter notebooks for exploration and analysis
├── src/                   # Source code
│   ├── pipeline/          # Data processing and modeling pipeline
│   ├── scripts/           # Helper scripts
│   └── main.py            # Main script to run the pipeline
├── .gitignore             # Files and folders ignored by Git
├── README.md              # Project description

🗃️ Dataset

• Tweets were scraped from Twitter using relevant keywords
• Brands: fore, kopi kenangan, point coffee, and tomoro
• All data was preprocessed and manually labeled into sentiment categories

🚀 Getting Started

1. Clone the Repository

git clone https://github.com/atrkhomeini/Sentiment-Analysis-of-4-Brand-Coffee-Shop-in-Indonesia-Using-CNN-BiLSTM.git
cd Sentiment-Analysis-of-4-Brand-Coffee-Shop-in-Indonesia-Using-CNN-BiLSTM

2. Install Dependencies

pip install -r requirements.txt

3. Run the Pipeline

python src/main.py

🧠 Model Architecture

• CNN layers extract local patterns and semantic cues (n-gram features)
• BiLSTM captures contextual dependencies from both forward and backward directions
• Dense layer with softmax classifies the output into sentiment categories

📈 Results Summary

• Achieved over 90% accuracy on validation data
• Fine-tuned using optimal CNN filter sizes, embedding dimensions, and dropout rates
• Output visualizations include accuracy/loss curves, confusion matrix, and sentiment word clouds

👤 Author

Ayat Tulloh Rahulloh Khomeini
Undergraduate in Information Systems, Institut Teknologi Sepuluh Nopember (ITS)
📫 Connect via LinkedIn or GitHub

📝 Citation

If you find this project useful in your research or development, please cite it as:

@misc{atrkhomeini2025sentiment,
  author       = {Ayat Tulloh Rahulloh Khomeini},
  title        = {Sentiment Analysis of Four Coffee Shop Brands in Indonesia Using CNN-BiLSTM},
  year         = {2025},
  url          = {https://github.com/atrkhomeini/Sentiment-Analysis-of-4-Brand-Coffee-Shop-in-Indonesia-Using-CNN-BiLSTM}
}

⚠️ Disclaimer

This repository is intended for educational and research purposes only. All brand names and trademarks are the property of their respective owners.

---

Built with ❤️ by Ayat Tulloh Rahulloh Khomeini