Fake News Detection

Author

Liantsoa Randrianasimbolarivelo
January 2025

Dataset Summary

The Fake News Classification Dataset contains over 45,000 unique English-language news articles. These articles are classified as true (1) or false (0), providing a valuable resource for fake news detection research using deep learning models.

Exploratory Data Analysis (EDA)

Dataset Structure

• Total records: 72,134 rows, 4 columns.

Data Types

Column Name	Data Type
Unnamed: 0	int64 (index)
title	object (text)
text	object (text)
label	int64 (binary classification)

Data Cleaning

• Removed the Unnamed: 0 column (irrelevant index).
• Checked for missing values:
  • title: 558 missing values.
  • text: 39 missing values.
  • label: 0 missing values.
• Checked for duplicated records: 8,416 duplicates found and removed.

Text Length Statistics

	Word Count	Character Count
Count	63,121	63,121
Mean	341.16	2451.03
Standard Deviation	369.81	2604.85
Min	2	11
25th Percentile	158	1131
Median	256	1857
75th Percentile	422	3055
Max	20,731	137,970

Target Variable Distribution

Most Frequent Words

Preprocessing

• Merged title and text columns into a single news variable.
• Removed punctuation and special characters.
• Converted text to lowercase.
• Removed stopwords (e.g., "the", "is", "and").
• Applied tokenization and lemmatization.
• Applied tokenizer transformation sequences to text.
• Applied padding to ensure uniform input size.

Modelization

The model uses a combination of word embeddings, Bidirectional Long Short-Term Memory (BiLSTM) layers, and regularization techniques to classify news articles as either fake or real. The model is implemented using TensorFlow and Keras.

Model Architecture

Embedding Layer

The first layer is an Embedding layer, which maps input words into dense vectors of fixed size.

• Vocabulary size: 10,000 words.
• Embedding dimensions: 20.
• Maximum input sequence length: 200 tokens.

Bidirectional LSTM Layers

Two BiLSTM layers capture contextual information from text in both forward and backward directions.

• First BiLSTM layer: 64 units, returns sequences.
• Second BiLSTM layer: 32 units, returns only final output.
• Both layers use L2 regularization (0.01) to prevent overfitting.

Batch Normalization and Dropout

• BatchNormalization: Normalizes activations for stable training.
• Dropout: 50% of the units are randomly dropped during training.

Dense Layers

• First dense layer: 16 units, ReLU activation.
• Output layer: 1 unit, Sigmoid activation (for binary classification).

Model Compilation

• Loss function: Binary cross-entropy.
• Optimizer: Adam optimizer.
• Evaluation metric: Accuracy.

Model Training and Performance

Training History

Epoch	Accuracy	Loss	Validation Accuracy
1	0.8935	1.1520	0.9505
2	0.9671	0.1210	0.9617
3	0.9802	0.0835	0.9598
4	0.9826	0.0782	0.9574
5	0.9869	0.0589	0.9607

Final Model Evaluation

• Test Accuracy: 95.98%
• Validation Accuracy: 96.06%