A Conditional Mutual Information-Based Approach for Robust Multi-Source Feature Selection in IoT Systems.

Area of Science:

• Computer Science
• Data Science
• Machine Learning

Background:

• High-dimensional data in Internet of Things (IoT) environments presents challenges like heterogeneity, redundancy, and noise.
• Effective feature selection is crucial for accurate analysis and model performance in such complex settings.

Purpose of the Study:

• To develop a feature selection method that balances classification performance, dimensionality reduction, and selection stability.
• To address the limitations of existing methods in handling noisy and redundant IoT data.

Main Methods:

• Proposes a residual-based conditional mutual information and feedback fusion (RCMF) feature-selection method.
• Introduces a residual-based indicator to quantify incremental discriminative information.
• Incorporates model-driven predictive contribution and stability scores with adaptive weight updates for fusion.

Main Results:

• The RCMF method jointly considers conditional discriminative information, task relevance, and selection consistency.
• Experiments on benchmark and IoT datasets demonstrate the method's rationality and effectiveness.
• Achieves a balance between classification performance, dimensionality reduction, and feature selection stability.

Conclusions:

• The proposed RCMF method offers a robust solution for feature selection in high-dimensional IoT data.
• It effectively enhances classification performance and selection stability in heterogeneous and noisy environments.
• Validated effectiveness across diverse datasets highlights its practical applicability.