Effective Model Update for Adaptive Classification of Text Streams in a Distributed Learning Environment.

Area of Science:

• Machine Learning
• Natural Language Processing
• Distributed Systems

Background:

• Adaptive classification models are crucial for analyzing dynamic text streams.
• Existing methods struggle with the scale and speed requirements of real-time data.
• Distributed learning environments offer potential for efficient model updates.

Purpose of the Study:

• To propose and evaluate dynamic model update strategies for adaptive text classification in distributed learning.
• To compare the effectiveness of entire model updates versus partial model updates.
• To assess the scalability of the proposed distributed learning architecture.

Main Methods:

• Developed two dynamic model update strategies: entire model update and partial model update.
• Applied strategies to Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), Bidirectional Long Short-Term Memory (Bi-LSTM), and BERT-based models.
• Conducted experiments on two real-world tweet streaming datasets.

Main Results:

• The entire model update strategy improved classification accuracy for pre-trained offline models.
• The partial model update strategy achieved comparable accuracy while significantly increasing learning speed.
• The distributed learning architecture demonstrated scalability, with reduced learning and inference times as worker nodes increased.

Conclusions:

• Both entire and partial model update strategies effectively enhance adaptive text classification in distributed environments.
• Partial model updates offer a compelling trade-off between accuracy and learning speed for time-sensitive applications.
• The proposed distributed learning framework is scalable and efficient for processing large-scale text streams.