Model Simplification of Deep Random Forest for Real-Time Applications of Various Sensor Data

Area of Science:

• Artificial Intelligence
• Machine Learning

Background:

• Deep Random Forests (DRF) offer high performance comparable to Deep Neural Networks (DNNs) without backpropagation.
• However, the complex structure of DRF, with numerous interconnected decision trees, hinders model interpretability and analysis.

Purpose of the Study:

• To propose a new method for simplifying black-box DRF models.
• To enhance the transparency and efficiency of DRF models through rule elimination.

Main Methods:

• Quantifying feature contributions and rule frequency within a trained DRF decision rule set.
• Implementing a rule elimination strategy based on measured feature contributions to remove unnecessary rules.

Main Results:

• The proposed method successfully simplified various DRF models and benchmark sensor datasets.
• Simplified DRF models demonstrated robust performance with significantly fewer parameters and rules.
• Comparison with compressed DNNs showed superior parameter compression and memory efficiency for the simplified DRF.

Conclusions:

• The developed rule elimination technique effectively simplifies DRF models.
• The simplified DRF models offer improved transparency, efficiency, and comparable classification accuracy to complex models.
• This approach provides a more interpretable and resource-efficient alternative to DNNs for complex tasks.