Constructing neural networks for multiclass-discretization based on information entropy

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Mining

Background:

• Previous entropy-based neural network architectures for discretization focused on two-class problems.
• Multiclass discretization methods often simplified inter-class relationships, leading to suboptimal performance.

Purpose of the Study:

• To develop an improved entropy-based neural network architecture for multiclass discretization.
• To enhance recognition rates by preserving inter-class relationships.

Main Methods:

• Proposed a novel method that considers entropy in a global sense for multiclass discretization.
• Maintained the inter-relationship among classes during neural network training.
• Enabled sharing of hidden nodes and layers among different classes.

Main Results:

• The proposed method achieved higher recognition rates compared to the traditional two-class-based approach.
• Allowed for efficient sharing of network components (hidden nodes and layers) across multiple classes.
• Addressed the limitations of previous methods that created independent subnetworks.

Conclusions:

• The global entropy-based approach is superior for multiclass discretization using neural networks.
• Preserving inter-class relationships and enabling node sharing leads to improved performance.
• This method offers a more effective strategy for complex discretization tasks.