Generating balanced learning and test sets for function approximation problems.

Area of Science:

• Machine Learning
• Data Mining
• Statistical Analysis

Background:

• Evaluating learning algorithms relies on partitioning data into learning and test sets.
• Improper partitioning can lead to unrepresentative sets and inaccurate algorithm performance conclusions.
• This issue is exacerbated with small datasets, increasing the risk of biased evaluations.

Purpose of the Study:

• To propose a deterministic data mining approach for creating balanced and representative learning and test sets.
• To ensure fair evaluation of learning algorithm accuracy and promote reproducible machine learning experiments.
• To address the limitations of random data partitioning, particularly for small datasets.

Main Methods:

• A novel deterministic data mining approach combining clustering and a nearest-neighbor distribution algorithm.
• Clustering is tailored for function approximation problems.
• Data is distributed into two sets within each cluster to maintain representativeness and balance.

Main Results:

• The proposed methodology generates balanced and representative learning and test sets.
• This approach enhances the reliability of evaluating machine learning algorithm accuracy.
• Experimental results demonstrate the effectiveness of the methodology through an ANOVA-based statistical study.

Conclusions:

• The deterministic data partitioning method offers a robust alternative to random splitting for function approximation.
• It leads to more accurate and reproducible assessments of learning algorithm performance.
• This technique is particularly valuable for small datasets where representativeness is critical.