Climate Misinformation Detector

This repository contains code for detecting climate misinformation using multiple machine learning and natural language processing (NLP) models.
The project evaluates traditional and transformer-based models on their ability to classify climate-related claims as supported or unsupported, while tracking energy use and carbon emissions for sustainable AI research.

---

Overview

Climate misinformation poses a major obstacle to informed action on environmental issues.
This project explores how well different machine learning approaches can identify misleading or false climate-related claims.
Models are trained and compared across performance, interpretability, and emissions efficiency.

Implemented Models

• Logistic Regression (custom + scikit-learn)
• Feedforward Neural Network (MLP) (scikit-learn)
• Transformer (custom implementation)
• Fine-tuned BERT (bert-base-uncased)

Each model is trained on a dataset of labeled climate claims and evaluated on standard classification metrics.

---

Results

Model	Accuracy	Recall	Fine Tuning CO₂ Emissions (g)	Training CO₂ Emissions (g)
Logistic Regression (scikit)	0.66	0.67	0.01	3
Neural Network (MLP)	0.68	0.75	0.01	3
Transformer (custom)	0.66	0.74	0.0	0.25
BERT-base Fine-tuned	0.66	0.67	0.01	650,000

---

Repository Structure

ninaiervin-climate_misinfo_detector/
│
├── BERT.py                       # Fine-tune a pretrained BERT model
├── BERT_eval.py                  # Evaluate fine-tuned BERT
│
├── transformer.py                # Train small Transformer from scratch
│
├── logistic_regression.py        # (Prototype) Custom logistic regression
├── scikit_logistic_regression.py # Logistic regression with sentence embeddings
├── logistic_regression_eval.py   # Evaluate logistic regression model
│
├── scikit_NN.py                  # Feedforward NN using scikit-learn
├── NN_eval.py                    # Evaluate trained NN model
│
├── exploring_data_layout.py      # Data preprocessing and splitting
│
├── log_reg_params_0.6818.joblib  # Saved logistic regression model (best accuracy)
│
└── data/                         # Expected folder for dataset (JSONL files)

---

Data Format

All scripts assume a dataset of climate claims in .jsonl format:

{
  "claim": "The Earth's climate has warmed over the past century.",
  "claim_label": "SUPPORTS"
}

Example expected files:

data/train_data.jsonl
data/dev_data.jsonl
data/test_data.jsonl

The helper script exploring_data_layout.py manages dataset loading and splitting.

---

Installation

Requirements

• Python 3.8+
• pip (Python package manager)

Install Dependencies

pip install torch transformers scikit-learn pandas numpy matplotlib joblib codecarbon

Optional:
To export results or plots, you may also install
pip install seaborn tqdm sentence-transformers

---

Data Exploration

python exploring_data_layout.py

Logistic Regression

Train:

python scikit_logistic_regression.py

Evaluate:

python logistic_regression_eval.py

Neural Network (MLP)

Train:

python scikit_NN.py -hi 128 -act relu -b 32 -lr 0.001

Evaluate:

python NN_eval.py

Transformer (from scratch)

Train:

python transformer.py --epochs 3 --batch_size 16

BERT Fine-Tuning

Train:

python BERT.py --epochs 3 --batch_size 16 --lr 2e-5 --output_dir ./bert_output

Evaluate:

python BERT_eval.py --output_dir ./bert_output

---

Sustainability Tracking

Each model’s energy usage and carbon emissions are tracked using codecarbon:

from codecarbon import EmissionsTracker
tracker = EmissionsTracker(project_name="ClimateMisinfo")
tracker.start()
# train model
tracker.stop()

Outputs include:

• Energy (kWh)
• CO₂ emissions (grams)
• Runtime duration (seconds)

---

Key Features

• End-to-end model comparison (from logistic regression to BERT)
• Consistent preprocessing and evaluation pipeline
• Integrated carbon tracking with codecarbon
• Modular, reproducible design for new datasets or models

---

🧑‍💻 Author

Nina Ervin
M.S. Computer Science, University of California San Diego
LinkedIn

Anuk Centellas
M.S. Computational Linguistics, University of Washington
LinkedIn

---

📜 License

This project is licensed under the MIT License.
See the LICENSE file for details.