Meta-Scaler: A Meta-Learning Framework for the Selection of Scaling Techniques.

Area of Science:

• Machine Learning
• Data Preprocessing
• Meta-Learning

Background:

• Dataset scaling, or normalization, is a critical preprocessing step in machine learning pipelines.
• Selecting the optimal scaling technique (ST) is dataset- and classifier-dependent, with no single ST universally outperforming others.
• Manual selection of the best ST via trial-and-error is often infeasible due to computational costs.

Purpose of the Study:

• To introduce Meta-scaler, a novel approach utilizing meta-learning (MtL) for automated selection of the best scaling technique.
• To develop meta-models that learn the relationship between dataset meta-features and the performance of classifiers under different STs.

Main Methods:

• Extracted meta-features from datasets.
• Trained meta-models to predict the best ST for a given dataset and classification algorithm.
• Evaluated Meta-scaler performance across 12 base classifiers and 300 datasets using five STs.

Main Results:

• Meta-scaler successfully identified optimal STs, leading to statistically significant performance improvements for 10 out of 12 base models compared to fixed ST choices.
• The proposed approach demonstrated superior performance over existing state-of-the-art MtL methods for ST selection.
• Experimental results validated the feasibility and effectiveness of the Meta-scaler approach.

Conclusions:

• Meta-scaler offers an effective automated solution for selecting the most suitable data scaling technique, enhancing machine learning model performance.
• The meta-learning framework provides a scalable and efficient alternative to manual ST selection.
• The study underscores the importance of adaptive ST selection for optimizing classification tasks.