A New Hybrid Deep Learning Method with Neutrosophic Sets for  Social Media Sentiment Analysis of the COVID-19 Vaccine

Tasneem Abdelrahman; Mohamed El-Rashidy; Mohamed Marie

Authors

Tasneem Abdelrahman Department of Computer Science, El Shorouk Academy, Cairo, Egypt.
Mohamed El-Rashidy Computer Science and Engineering Department, Faculty of Electronic Engineering, Menoufia University, Menoufia, Egypt.
Mohamed Marie Information Systems Department, Faculty of Computers and Artificial Intelligence, Helwan University, Cairo, Egypt.

Keywords:

Single Valued Neutrosophic Set; Decision Making; COVID-19; Vaccine Sentiment Analysis; Deep Learning, Convolutional Neural Networks (CNN); Long Short-Term Memory (LSTM); TF-IDF; Social Media Analytics

Abstract

The COVID-19 pandemic has launched historic public debates on vaccines, and more significantly,
on social media posts regarding vaccines. In this study, we investigate how the public thinks about COVID
19 vaccinations using state-of-the-art machine learning and deep learning techniques. Our proposed hybrid
CNN + LSTM model achieved the best performance (94.7% accuracy, 95% precision, 95% recall, and 95% F1
score) among individual and traditional deep learning models. Deep learning techniques outperform the
conventional classifiers in a comparative evaluation of different models, i.e., Attention-based GRU, Bi-LSTM,
Bi-GRU, Hybrid Bi-LSTM + GRU, and Logistic Regression (Unigram, Bigram, Trigram). Attention Layer on
GRU and Attention Layer on CNN + Bi-LSTM models also performed well, registering accuracy values of
94.62% and 94.37%, respectively. Also, we propose a decision-making model to select the best model under
different evaluation matrices. Seven models are evaluated under four evaluation matrices. We use the
neutrosophic sets to deal with uncertainty in the evaluation process. The RAM method is used to select the
best model. The dataset included over 200,000 tweets that had been preprocessed for sentiment categorization
using VADER (Valence Aware Dictionary and Sentiment Reasoner) and Text Blob, feature extraction, and TF
IDF. Temporal sentiment analysis also found that while negative sentiment experienced transient peaks in
reaction to negative vaccination narratives and disinformation, positive sentiment rose consistently over time
and was tied to vaccine rollout efforts and public health messages. The results show how hybrid deep learning
models may be used in sentiment analysis, and they give public health officials useful information to combat
false information and enhance the targeting of vaccination campaigns.