3HAN_pytorch

This repository contains 3HAN: A Hierarchical Attention Network for Fake News Detection. Official Paper link: https://link.springer.com/chapter/10.1007/978-3-319-70096-0_59

Model Introduction

3HAN utilizes three attention levels, focusing on words, sentences, and the headline, to construct a news vector: an effective representation of an input news article. It processes the article in a hierarchical bottom-up manner. Since the headline is a distinguishing feature of fake news and only a few words and sentences in an article carry more significance than others, 3HAN assigns differential importance to various parts of the article through its three layers of attention. Basically: Word Level Attention -> Sentence Level Attention -> Title Level Attention -> Final News Vector. BidirectionalGRU is used for sequence encoding.

Datasets used

I used the following datasets from Kaggle: https://www.kaggle.com/datasets/emineyetm/fake-news-detection-datasets

Train Instructions

If you want to change the parameters, check out Train.py-> Class Options where you can also add the path to your data.

To run the code, simply go to the terminal and type:

python Train.py

Results

Breakdown of the Confusion Matrix ---->

True Positives (TP): 115,682 (bottom right)

True Negatives (TN): 114,152 (top left)

False Positives (FP): 5,348 (top right)

False Negatives (FN): 4,818 (bottom left)

Accuracy: The overall accuracy of the model is (TP + TN) / Total = (115,682 + 114,152) / (115,682 + 114,152 + 5,348 + 4,818) ≈ 0.957. SO about 95.7% of the predictions were correct.

Precision (True Class): TP / (TP + FP) = 115,682 / (115,682 + 5,348) ≈ 0.956

Recall (True Class):TP / (TP + FN) = 115,682 / (115,682 + 4,818) ≈ 0.960

There is a good balance between false positives and false negatives, so the model does not heavily favor one class over the other.