Multitree Genetic Programming With Feature-Based Transfer Learning for Symbolic Regression on Incomplete Data.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Data incompleteness is a significant challenge in real-world machine learning, particularly impacting symbolic regression (SR) algorithms.
• Existing SR methods struggle with limited or missing data, hindering their learning capabilities.
• Transfer learning (TL) offers a promising solution for knowledge transfer but remains underexplored in SR contexts.

Purpose of the Study:

• To propose and evaluate a novel multitree genetic programming-based transfer learning (TL) method for symbolic regression (SR) in the presence of data incompleteness.
• To address the challenge of knowledge transfer from complete source domains (SDs) to incomplete target domains (TDs).
• To enhance the performance and efficiency of SR algorithms when dealing with missing data.

Main Methods:

• A multitree genetic programming approach is employed for transfer learning.
• Features are transformed from complete source domains to incomplete target domains.
• An integrated feature selection mechanism is utilized to streamline the transformation process by removing redundant features.

Main Results:

• The proposed TL method demonstrates effectiveness in handling missing values across both real-world and synthetic SR tasks.
• The method shows improved training efficiency compared to existing TL approaches.
• Significant reductions in regression error were observed: an average of over 2.58% on heterogeneous domains and 4% on homogeneous domains compared to state-of-the-art methods.

Conclusions:

• The developed multitree genetic programming-based TL method successfully addresses data incompleteness in symbolic regression.
• The integration of feature selection enhances the practicality and efficiency of knowledge transfer.
• This approach offers a robust solution for improving SR performance in data-scarce and incomplete scenarios.