Interpretability Diversity for Decision-Tree-Initialized Dendritic Neuron Model Ensemble.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computational Science

Background:

• Classifier ensembles require accurate and diverse base classifiers for optimal performance.
• Existing diversity metrics lack a standardized definition and measurement approach.
• Interpretability is a crucial but often overlooked aspect of classifier diversity.

Purpose of the Study:

• To propose a novel metric, Learners' Interpretability Diversity (LID), for quantifying the diversity of interpretable machine learners.
• To develop a LID-based classifier ensemble method that leverages interpretability for improved diversity.
• To demonstrate the effectiveness of the proposed ensemble using decision-tree-initialized dendritic neuron models (DDNM).

Main Methods:

• Introduced Learners' Interpretability Diversity (LID) as a quantitative measure for interpretable learner diversity.
• Developed a classifier ensemble strategy based on the proposed LID metric.
• Utilized decision-tree-initialized dendritic neuron models (DDNM) as base learners for ensemble construction.
• Evaluated the ensemble's performance on seven benchmark datasets.

Main Results:

• The LID-based DDNM ensemble demonstrated superior performance compared to popular classifier ensembles.
• The proposed method achieved significant improvements in both accuracy and computational efficiency.
• A random-forest-initialized dendritic neuron model (RDNM) combined with LID emerged as a highly effective representative of the DDNM ensemble.

Conclusions:

• Learners' Interpretability Diversity (LID) offers a novel and effective approach to measuring and utilizing classifier diversity.
• LID-based ensembles provide a promising direction for developing more accurate and computationally efficient machine learning models.
• The interpretability of base learners can be effectively leveraged to enhance ensemble performance.