Controlling Complexity and Accuracy of Classification Decision Tree Extracted from Trained Artificial Neural Network

2019 60th International Scientific Conference on Information Technology and Management Science of Riga Technical University (ITMS 2019): Proceedings 2019
Andrey Bondarenko

There is growing number of publications devoted to knowledge extraction from fully connected feed-forward artificial neural networks. Although there are not many publications covering ways allowing to control extracted knowledge complexity and precision. The higher complexity is, the higher accuracy can be gained. But in case ANN should be validated by domain expert or just be explainable it should be simple enough – this will lower accuracy of extracted knowledge. The current paper explores influence of parameters used for ANN pruning and neurons outputs discretization and clustering onto accuracy of extracted classification decision tree. Hence reader is presented with experimental validation of effects produced by variation in parameters combination.

Keywords
Knowledge extraction, neural networks, classification
DOI
10.1109/ITMS47855.2019.8940739
Hyperlink
https://ieeexplore.ieee.org/document/8940739

Bondarenko, A. Controlling Complexity and Accuracy of Classification Decision Tree Extracted from Trained Artificial Neural Network. In: 2019 60th International Scientific Conference on Information Technology and Management Science of Riga Technical University (ITMS 2019): Proceedings, Latvia, Riga, 10-11 October, 2019. Piscataway: IEEE, 2019, pp.1-6. ISBN 978-1-7281-5710-8. e-ISBN 978-1-7281-5709-2. Available from: doi:10.1109/ITMS47855.2019.8940739

Publication language
English (en)