Algorithm Recommendation and Performance Prediction Using Meta-Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• The proliferation of machine learning algorithms and parameters has increased model-finding potential but also training complexity.
• Growing dataset sizes exacerbate computational costs and time for traditional exhaustive search methods, particularly in data streaming.
• Existing automated machine learning (AutoML) approaches can be computationally intensive.

Purpose of the Study:

• To develop a meta-learning approach for predicting machine learning model performance indicators.
• To recommend optimal algorithm and configuration pairings for specific machine learning problems.
• To address the computational challenges of model training in large-scale and streaming data environments.

Main Methods:

• A meta-learning framework was employed to analyze algorithm performance across diverse datasets.
• The approach focuses on predicting key performance indicators (KPIs) rather than exhaustive model training.
• It generates recommendations for algorithm and hyperparameter selection.

Main Results:

• The proposed meta-learning method achieved up to 130x speed improvement compared to a state-of-the-art AutoML method.
• The trade-off in average model quality was minimal, with only a 4% decrease.
• The approach demonstrates significant efficiency gains for model selection and training.

Conclusions:

• Meta-learning offers a computationally efficient alternative for machine learning model selection and training.
• The method is particularly advantageous for time-sensitive applications like big data streaming where rapid model updates are necessary.
• This approach enables faster iteration and deployment of machine learning models in resource-constrained or dynamic environments.