[Paper Review] Combination Of Convolution Neural Networks And Deep Neural Networks For Fake News Detection

Zainab A. Jawad, Ahmed J. Obaid|arXiv (Cornell University)|Oct 15, 2022

Misinformation and Its Impacts4 citations

TL;DR

This paper proposes a hybrid deep learning model combining Convolutional Neural Networks (CNN) and Deep Neural Networks (DNN) for fake news detection using the FNC-1 dataset. It employs NLP techniques, soft cosine similarity, and feature reduction via the elbow method, achieving 84.6% accuracy—ranking second among competitive submissions—except for the 'disagree' category.

ABSTRACT

Nowadays, People prefer to follow the latest news on social media, as it is cheap, easily accessible, and quickly disseminated. However, it can spread fake or unreliable, low-quality news that intentionally contains false information. The spread of fake news can have a negative effect on people and society. Given the seriousness of such a problem, researchers did their best to identify patterns and characteristics that fake news may exhibit to design a system that can detect fake news before publishing. In this paper, we have described the Fake News Challenge stage #1 (FNC-1) dataset and given an overview of the competitive attempts to build a fake news detection system using the FNC-1 dataset. The proposed model was evaluated with the FNC-1 dataset. A competitive dataset is considered an open problem and a challenge worldwide. This system's procedure implies processing the text in the headline and body text columns with different natural language processing techniques. After that, the extracted features are reduced using the elbow truncated method, finding the similarity between each pair using the soft cosine similarity method. The new feature is entered into CNN and DNN deep learning approaches. The proposed system detects all the categories with high accuracy except the disagree category. As a result, the system achieves up to 84.6 % accuracy, classifying it as the second ranking based on other competitive studies regarding this dataset.

Motivation & Objective

• To address the growing challenge of fake news dissemination on social media platforms.
• To develop a robust, automated system capable of detecting fake news before publication.
• To improve detection accuracy by combining CNN and DNN architectures with advanced feature engineering.
• To evaluate the model on the benchmark FNC-1 dataset and compare performance with existing approaches.
• To identify limitations in detecting the 'disagree' category, which remains a challenge.

Proposed method

• Text from headlines and body texts is preprocessed using natural language processing techniques.
• Features are extracted and reduced using the elbow truncated method to optimize dimensionality.
• Soft cosine similarity is applied to measure similarity between feature pairs.
• The reduced features are fed into a CNN for local pattern extraction and a DNN for high-level representation learning.
• The final classification is performed by combining the outputs of CNN and DNN components.
• The model is trained and evaluated on the FNC-1 dataset using standard metrics.

Experimental results

Research questions

• RQ1 Can a hybrid CNN-DNN architecture effectively detect fake news using textual features from headlines and body text?
• RQ2 How does feature reduction via the elbow method impact model performance in fake news detection?
• RQ3 To what extent does soft cosine similarity improve feature representation for fake news classification?
• RQ4 Why does the model perform poorly on the 'disagree' category compared to other classes?
• RQ5 How does the proposed model compare in accuracy to other state-of-the-art systems on the FNC-1 benchmark?

Key findings

• The proposed model achieves an overall accuracy of 84.6% on the FNC-1 dataset, ranking second among competing systems.
• The model performs well across all fake news categories except the 'disagree' category, which shows lower detection accuracy.
• Feature reduction using the elbow method effectively balances model complexity and performance.
• Soft cosine similarity contributes to improved feature representation by capturing semantic relationships.
• The integration of CNN and DNN enables effective learning of both local and global text patterns.
• The model demonstrates strong generalization on the benchmark dataset, indicating robustness to diverse fake news patterns.