README.md

metadata

language:
  - en
license: cc-by-nc-sa-4.0
size_categories:
  - 1M<n<10M
task_categories:
  - graph-ml
  - text-classification
pretty_name: TAGFN
configs:
  - config_name: politifact
    data_files:
      - politifact/raw_text/*.parquet
    default: true
  - config_name: gossipcop
    data_files:
      - gossipcop/raw_text/*.parquet
  - config_name: fakeddit
    data_files:
      - fakeddit/raw_text/*.parquet
tags:
  - social
  - news
  - outlier
  - graph
  - misinformation
  - classification
  - detection

TAGFN: A Text-Attributed Graph Dataset for Fake News Detection in the Age of LLMs

Paper | Code

TAGFN is a large-scale, real-world text-attributed graph dataset specifically designed for outlier detection, particularly in the context of fake news detection. It addresses the scarcity of large-scale, realistic, and well-annotated datasets for evaluating both traditional and Large Language Model (LLM)-based graph outlier detection methods. This dataset also facilitates the fine-tuning of LLMs for developing misinformation detection capabilities, serving as a valuable resource for advancing robust graph-based outlier detection and trustworthy AI.

Sample Usage

To reproduce experiments, navigate to the repository and run:

bash run.sh

To host an LLM server locally with vllm, refer to the commands provided in vllm.sh.