Federal learning edge network based sentiment analysis combating global COVID-19.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Sentiment analysis is crucial for understanding public opinion on COVID-19.
• Deep learning models for sentiment analysis face challenges with limited and distributed datasets.
• Data privacy concerns hinder the use of large social media datasets.

Purpose of the Study:

• To propose a novel federal learning framework integrating Bidirectional Encoder Representations from Transformer (BERT) and multi-scale convolutional neural networks (MSCNN) for COVID-19 sentiment analysis.
• To overcome data scarcity and privacy issues in sentiment analysis of COVID-19 related web data.
• To improve the efficiency and performance of sentiment analysis models.

Main Methods:

• Developed a federal learning framework (Fed-BERT-MSCNN) combining BERT and MSCNN.
• Implemented a central server and local machines for distributed training on social platform datasets.
• Utilized edge networks for secure parameter communication and weighted averaging.

Main Results:

• The Fed-BERT-MSCNN model demonstrated superior performance compared to existing sentiment analysis models.
• The proposed framework effectively addressed data limitations and ensured data privacy during training.
• Improved communication efficiency was observed within the federal network.

Conclusions:

• Fed-BERT-MSCNN offers a robust solution for sentiment analysis on sensitive, distributed datasets.
• Federal learning with BERT and MSCNN is a promising approach for public health data analysis.
• The model enhances sentiment analysis accuracy and privacy for COVID-19 research.