A model for single and multiple knowledge based networks

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Biology

Background:

• Neural networks often function as black boxes, hindering the explanation of their decisions.
• Existing research on rule extraction primarily focuses on single neural networks, with limited investigation into combined networks.

Purpose of the Study:

• To develop a method for translating symbolic rules into a Discretized Interpretable Multi-Layer Perceptron (DIMLP) model.
• To enable rule extraction from single and multiple combined neural networks.
• To enhance the interpretability of neural network responses.

Main Methods:

• Characterizing discriminant hyperplane frontiers for rule extraction.
• Translating symbolic rules into the DIMLP model.
• Applying single DIMLP networks and ensembles to medical diagnosis and prognosis datasets.

Main Results:

• Unordered rules extracted in polynomial time with 100% matching accuracy on training data.
• DIMLP models demonstrated comparable accuracy to standard Multi-Layer Perceptrons (MLP) across 17 medical datasets.
• DIMLP networks significantly outperformed the CN2 algorithm on eight problems, particularly in non-Hodgkin lymphoma diagnosis (96.1% accuracy).

Conclusions:

• The DIMLP approach effectively extracts symbolic rules from neural networks, enhancing model interpretability.
• DIMLP networks offer a powerful and accurate alternative for medical diagnosis and prognosis tasks.
• Ensembles of DIMLP networks show superior performance in complex classification problems like lymphoma diagnosis.