Belief Entropy Tree and Random Forest: Learning from Data with Continuous Attributes and Evidential Labels

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Mining

Background:

• Decision trees are established machine learning algorithms for classification.
• Handling uncertainty in data labels is a significant challenge in pattern recognition.
• Existing methods often require data discretization for continuous attributes.

Purpose of the Study:

• To propose a new decision tree method that incorporates belief functions to manage label uncertainty.
• To extend this method into a random forest for improved classification performance.
• To enable direct handling of continuous attributes without discretization.

Main Methods:

• A novel decision tree algorithm utilizing belief entropy for attribute selection.
• Integration of Gaussian mixture models for direct processing of continuous data.
• Construction of a random forest ensemble using the proposed decision trees.
• Exploration of various prediction combination strategies for the random forest.

Main Results:

• The proposed decision tree and random forest method demonstrate good classification accuracy on UCI datasets.
• The method excels particularly when dealing with datasets exhibiting significant label uncertainty.
• Direct handling of continuous attributes avoids preprocessing steps like discretization, simplifying the workflow.

Conclusions:

• The belief entropy-based decision tree and random forest offer a robust approach for classification tasks with uncertain labels.
• This method provides a valuable alternative for analyzing complex datasets, especially those with inherent ambiguity.
• The approach enhances classification performance and efficiency by directly managing continuous data and uncertainty.