An Empirical Investigation Into Deep and Shallow Rule Learning

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Mining

Background:

• Inductive rule learning is a traditional machine learning paradigm.
• Current state-of-the-art learners create descriptions directly relating input features to target concepts, often using disjunctive normal form (DNF).
• While DNF is logically sufficient, deeper, more structured representations (deep theories) might offer learning advantages, analogous to deep neural networks.

Purpose of the Study:

• To empirically compare the performance of deep and shallow rule sets.
• To investigate the potential of deep rule structures in machine learning.
• To provide evidence for the value of developing deep rule learning algorithms.

Main Methods:

• Developed and optimized deep and shallow rule sets using a uniform mini-batch optimization algorithm.
• Conducted comparative experiments on both artificial and real-world benchmark datasets.
• Evaluated performance metrics of deep versus shallow rule networks.

Main Results:

• Deep rule networks consistently outperformed their shallow counterparts across diverse datasets.
• The findings suggest that deeper, more structured rule representations are more effective.
• The optimization approach facilitated a direct comparison between deep and shallow rule learning.

Conclusions:

• Deep rule structures demonstrate superior performance compared to shallow rule sets.
• The study supports the hypothesis that deep theories can be more effectively learned than shallow ones.
• Further research and development efforts in deep rule learning are warranted.