Greedy rule generation from discrete data and its use in neural network rule extraction

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Mining

Background:

• Classification rule generation is crucial for interpretable AI.
• Extracting rules from complex models like neural networks remains a challenge.
• Existing methods may lack efficiency or conciseness in rule sets.

Purpose of the Study:

• To introduce a novel Greedy Rule Generation (GRG) algorithm.
• To develop an efficient method for generating classification rules from discrete data.
• To extract accurate and concise rules from trained and pruned neural networks.

Main Methods:

• The proposed GRG algorithm iteratively selects the best rule based on coverage and attribute count.
• Classification rules were extracted from neural networks using a standard decompositional approach.
• The GRG algorithm was applied to discretized hidden unit activation values of trained neural networks.

Main Results:

• The GRG method produced accurate and concise rule sets when applied to neural network rule extraction.
• Experimental results demonstrated the effectiveness of GRG for rule generation on discrete medical datasets.
• The algorithm's greedy approach optimizes rule selection criteria.

Conclusions:

• The GRG algorithm offers an effective and efficient approach for classification rule generation.
• This method successfully bridges the gap between neural network performance and rule interpretability.
• GRG shows significant potential for applications in domains requiring clear, data-driven rules, such as healthcare.