Classification Tree Extraction from Trained Artificial Neural Networks
Procedia Computer Science 2016
Andrejs Bondarenko, Ludmila Aleksejeva, Vilens Jumutcs, Arkādijs Borisovs

Recent advances in neural networks design and training provoked 2nd artificial neural networks (ANN) renaissance. In many cases classification decision made by trained fully connected neural nets is better than one acquired by models like C4.5 or C5.01,2. But in contrast to decision trees ANN models are “black boxes”, i.e. it is impossible to understand how classification decision is made. In many areas it is critical and even obligate to understand how model performs classification thus rendering ANN usage as obsolete. Recently some researchers have proposed and described separate steps that would allow to extract knowledge from trained multi-layered fully connected sigmoidal neural network. This process involves several steps such as trained network training, pruning and knowledge extraction. Current paper provides overview of all of mentioned steps, as well as describes how knowledge extraction system can be built. We describe our Neural Network Knowledge eXtraction (NNKX) system and provide experimental results of rules extraction from trained multi-layered feed-forward sigmoidal artificial neural network in the form of binary classification decision trees. Results suggest that extracted decision trees have good classification accuracy and sizes comparable to C4.5 trees and even overcoming them in some cases. Thus proposed system can be successfully applied to better understand and validate ANN models. We provide link to source code repository with the implementation of described system.


Atslēgas vārdi
Neural networks; Knowledge extraction; Classification decision tree
DOI
10.1016/j.procs.2017.01.172
Hipersaite
http://www.sciencedirect.com/science/article/pii/S1877050917301734?via%3Dihub

Bondarenko, A., Aleksejeva, L., Jumutcs, V., Borisovs, A. Classification Tree Extraction from Trained Artificial Neural Networks. Procedia Computer Science, 2016, Vol.104, 556.-563.lpp. ISSN 1877-0509. Pieejams: doi:10.1016/j.procs.2017.01.172

Publikācijas valoda
English (en)
RTU Zinātniskā bibliotēka.
E-pasts: uzzinas@rtu.lv; Tālr: +371 28399196