Self-Adaptive Induction of Regression Trees.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Data streams often exhibit complex dynamics, including gradual and abrupt function drift, regional changes, and noise.
• Existing regression methods require careful configuration to adapt to these evolving data stream characteristics.
• Handling both symbolic and numeric attributes in dynamic environments remains a challenge.

Purpose of the Study:

• To introduce SAIRT, a novel algorithm for the incremental construction of binary regression trees.
• To enable adaptive model induction in data streams with unknown and changing dynamics.
• To improve the robustness and performance of regression models in dynamic environments.

Main Methods:

• SAIRT incrementally constructs binary regression trees, adapting the model to data stream dynamics.
• The algorithm automatically adjusts internal parameters and model structure to capture new patterns.
• It monitors node usefulness and employs local windows at tree leaves to manage data, including forgetting irrelevant examples.

Main Results:

• SAIRT effectively handles gradual and abrupt function drift, regional changes, noise, and virtual drift.
• The algorithm demonstrates superior performance compared to current methods across various data stream dynamics.
• It achieves better results with varying speeds of change, noise levels, sampling distributions, and function alterations.

Conclusions:

• SAIRT offers a robust solution for incremental regression on data streams with complex, unknown dynamics.
• The adaptive nature of SAIRT reduces the need for manual parameter tuning in dynamic environments.
• SAIRT advances the state-of-the-art in handling evolving data streams for regression tasks.