Statistical Mechanics of On-Line Learning Under Concept Drift

Area of Science:

• Machine Learning
• Statistical Physics
• Non-stationary Environments

Background:

• Online machine learning models often assume stationary environments, which rarely holds true in real-world applications.
• Concept drift, where the target variable changes over time, poses significant challenges for model training and performance.
• Existing methods for analyzing training dynamics are often limited to stationary conditions.

Purpose of the Study:

• To develop and exemplify a modeling framework for investigating online machine learning in non-stationary environments.
• To extend statistical physics methods for analyzing training dynamics under concept drift.
• To evaluate the performance of Learning Vector Quantization (LVQ) and layered neural networks in the presence of continuous target changes.

Main Methods:

• Developed a modeling framework for online machine learning in non-stationary settings.
• Applied prototype-based Learning Vector Quantization (LVQ) for classification tasks.
• Studied layered neural networks with sigmoidal activations for regression tasks.
• Extended statistical physics methods to analyze training dynamics with concept drift.

Main Results:

• Demonstrated LVQ's capability to track evolving classification schemes under drift to a significant extent.
• Showed that concept drift can lead to the persistence of suboptimal plateau states in gradient-based neural network training for regression.
• Presented first results for stochastic drift processes in both classification and regression.

Conclusions:

• The proposed framework effectively models online learning in non-stationary environments, offering insights into model behavior under concept drift.
• LVQ shows resilience in adapting to changing classification tasks, highlighting its potential for dynamic environments.
• Gradient-based neural network training for regression is susceptible to concept drift, potentially hindering convergence to optimal solutions.