Measuring the Effectiveness of Adaptive Random Forest for Handling Concept Drift in Big Data Streams

Area of Science:

• Machine Learning
• Data Science
• Stream Data Processing

Background:

• Big data and stream data processing present real-time challenges.
• Concept drift significantly impacts machine learning models on data streams.
• Adaptive Random Forest (ARF) is a promising stream learning algorithm for handling data distribution changes.

Purpose of the Study:

• To propose and evaluate a resampling mechanism for enhancing the efficiency of streaming algorithms like ARF.
• To introduce a measure of resampling effectiveness (ρ) that balances accuracy and execution time.
• To optimize ARF performance by tuning the Poisson distribution parameter (λ) for resampling.

Main Methods:

• Empirical selection of the Poisson distribution parameter (λ) using six synthetic datasets with varying drift types.
• Comparison of standard ARF with tuned variations based on resampling effectiveness (ρ).
• Validation of the proposed enhancement method on three real-world case studies: Amazon reviews, Arabic hotel reviews, and COVID-19 tweet sentiment analysis.

Main Results:

• The proposed resampling enhancement method demonstrated considerable improvement in ARF performance across various scenarios.
• Tuning the Poisson distribution parameter (λ) effectively balances accuracy and execution time for online learning.
• The method proved effective in processing large datasets and handling different types of concept drift.

Conclusions:

• Optimizing resampling strategies is key to improving the efficiency of stream learning algorithms.
• The proposed method offers a practical enhancement for ARF, leading to better performance in big data applications.
• This approach is effective for real-time processing of diverse data streams, including sentiment analysis.