Distributed Selection of Continuous Features in Multilabel Classification Using Mutual Information.

Area of Science:

• Machine Learning
• Data Science
• Big Data Analytics

Background:

• Multilabel learning presents scalability challenges for large datasets.
• Feature selection is crucial for improving multilabel accuracy and managing high dimensionality.
• Existing methods struggle with the complexity of continuous features in distributed environments.

Purpose of the Study:

• To propose a distributed model for mutual information (MI) adaptation on continuous features and multiple labels.
• To develop efficient and scalable feature selection methods for large-scale multilabel data.
• To address the curse of dimensionality in distributed computing.

Main Methods:

• Implemented a distributed model for mutual information adaptation on Apache Spark.
• Developed two MI-based feature selection approaches: MI maximization and minimum redundancy-maximum relevance (MIM).
• Evaluated methods on 10 datasets using 12 metrics.

Main Results:

• The proposed distributed methods outperform existing reference methods for multilabel feature selection.
• The MIM approach demonstrated significant reductions in runtime, orders of magnitude faster.
• Statistical analysis validated the superior performance and efficiency of the developed methods.

Conclusions:

• The novel distributed approaches effectively enhance multilabel feature selection accuracy and scalability.
• Apache Spark provides a robust platform for efficient distributed multilabel feature selection.
• These methods offer a significant advancement for handling complex, high-dimensional multilabel data in big data scenarios.