---
title: IndoHoaxDetector
emoji: 🌖
colorFrom: pink
colorTo: indigo
sdk: gradio
sdk_version: 5.49.1
app_file: app.py
pinned: false
license: mit
short_description: Indonesian News Detector
tags: ["nlp", "text-classification", "indonesian", "hoax-detection", "machine-learning", "fact-checking", "transformers", "tfidf"]
---

# IndoHoaxDetector

A comprehensive machine learning system for detecting hoax-style news articles in Indonesian language. This project provides multiple models trained on linguistic features of Indonesian news to identify articles written in a style typical of hoaxes or fake news. The system analyzes writing patterns, sensationalism, and other stylistic indicators rather than checking factual accuracy.

## Features

- **Multiple Models**: Logistic Regression, Linear SVM, Random Forest, Naive Bayes, and IndoBERT
- **High Accuracy**: Best model (IndoBERT) achieves 99.89% accuracy on test data
- **Web Interface**: Interactive Gradio app for easy prediction
- **Lightweight Options**: TF-IDF based models for fast inference
- **Transformer Support**: IndoBERT for state-of-the-art performance

## Model Comparison

All models were trained on 62,972 Indonesian news articles and evaluated on 12,595 held-out test samples:

| Model                  | Accuracy | Precision (Macro) | Recall (Macro) | F1 (Macro) | Type |
|------------------------|----------|-------------------|----------------|------------|------|
| **IndoBERT**          | **0.9989** | **0.9989**      | **0.9989**   | **0.9989** | Transformer |
| Linear SVM            | 0.9819  | 0.9820          | 0.9817       | 0.9818    | TF-IDF + SVM |
| Logistic Regression   | 0.9782  | 0.9787          | 0.9777       | 0.9781    | TF-IDF + LR |
| Random Forest         | 0.9765  | 0.9768          | 0.9760       | 0.9764    | TF-IDF + RF |
| Multinomial Naive Bayes| 0.9398 | 0.9414          | 0.9381       | 0.9393    | TF-IDF + NB |

## Model Details

### IndoBERT (Recommended)
- **Type**: Fine-tuned IndoBenchmark/indobert-base-p1
- **Training**: 3 epochs, batch_size=16, learning_rate=2e-5, max_length=128
- **Accuracy**: 99.89%
- **Best For**: Highest accuracy, contextual understanding
- **Requirements**: transformers, torch (GPU recommended)

### TF-IDF Based Models
- **Vectorizer**: TF-IDF (unigrams, max_features=5000, sublinear_tf=True)
- **Preprocessing**: Lowercase, URL removal, punctuation, Indonesian stopwords, Sastrawi stemming
- **Logistic Regression**: L2 regularization, max_iter=1000 (97.82% accuracy)
- **Linear SVM**: C=1.0, max_iter=10000 (98.19% accuracy)
- **Random Forest**: 100 trees, random_state=42 (97.65% accuracy)
- **Naive Bayes**: Multinomial, alpha=1.0 (93.98% accuracy)

## Usage

### Web Interface
Visit the Hugging Face Space to use the interactive app:
1. Select a model from the dropdown
2. Enter Indonesian news text
3. Click "Detect"
4. View the prediction and confidence

### Local Usage
To run locally:

```bash
pip install gradio scikit-learn transformers torch
python app.py
```

### API Usage
Load and use models programmatically:

```python
import pickle
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# For TF-IDF models (Logistic Regression, SVM, RF, NB)
with open('logreg_model.pkl', 'rb') as f:
    model = pickle.load(f)

# Load vectorizer
with open('tfidf_vectorizer.pkl', 'rb') as f:
    vectorizer = pickle.load(f)

text = "Your Indonesian news text here"
X = vectorizer.transform([text])
prediction = model.predict(X)  # 0 for FAKTA, 1 for HOAX

# For IndoBERT
tokenizer = AutoTokenizer.from_pretrained('indobert_model')
model = AutoModelForSequenceClassification.from_pretrained('indobert_model')

inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)
with torch.no_grad():
    outputs = model(**inputs)
    prediction = torch.argmax(outputs.logits, dim=-1).item()
```

## Training Data

- **Source**: TurnBackHoax fact-checks, Kompas/CekFakta.com articles
- **Size**: 62,972 labeled articles (balanced HOAX/FAKTA)
- **Preprocessing**: Indonesian text normalization, stemming, stopword removal
- **Split**: 80% train, 20% test (stratified)

## Limitations

- **Stylistic Analysis Only**: Detects hoax-like writing style, not factual accuracy
- **Language Specific**: Trained on Indonesian news patterns
- **Domain Shift**: May perform differently on social media vs. formal news
- **False Positives/Negatives**: Sensational legitimate news may be flagged as hoax
- **Transformer Requirements**: IndoBERT needs GPU for training/inference

## Ethical Considerations

- **Responsible Use**: Augment human fact-checkers, not replace them
- **Bias Awareness**: Training data may reflect source biases
- **Transparency**: All models and code are open-source
- **Misuse Prevention**: Clearly labeled as stylistic detector, not truth verifier

## Contributing

Contributions welcome! Open issues or PRs for improvements.

## License

MIT License - see LICENSE file for details

Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference