Comprehensible and transparent rule extraction using neural network

• Published in: Multimedia tools and applications 2024-02, Vol.83 (28), p.71055-71070
• Main Authors: Biswas, Saroj Kr, Bhattacharya, Arijit, Duttachoudhury, Abhinaba, Chakraborty, Manomita, Das, Akhil Kumar
• Format: Article
• Language: English
• Subjects: Accuracy; Algorithms; Classification; Computer Communication Networks; Computer Science; Data mining; Data Structures and Information Theory; Datasets; Machine learning; Multimedia Information Systems; Neural networks; Special Purpose and Application-Based Systems
• ISSN: 1573-7721; 1380-7501; 1573-7721
• DOI: 10.1007/s11042-024-18254-4

In data mining and machine learning communities, Neural Network (NN) is a popular classification method. On extremely unbalanced and complicated datasets, NN may achieve excellent classification accuracy. However, one disadvantage of NN is its inability to explain its reasoning process, which restricts its use in numerous sectors that need clear conclusions as well as high accuracy. To address this issue, rule-extraction mechanisms exist that extract intelligible classification-rules from NN and turn them into a white box. Attribute or network pruning, dealing with diverse attribute types, rule pruning, and dealing with class overlapping difficulties are all significant components or portions of many existing rule extraction methods, and present strategies to deal with these aspects are insufficiently successful. As a result, this study offers a rule extraction approach named “Comprehensible and Transparent Rule Extraction Using Neural Network”-CTRENN to address the aforementioned shortcomings and transform NN into a white box with high accuracy and better explain-ability. The suggested CTRENN is an expansion of the state-of-art Rule Extraction from Neural Network Using Classified and Misclassified Data technique (RxNCM). The CTRENN augments the RxNCM with a floating sequential search for feature and rule selection to improve feature and rule selection. CTRENN also distinguishes between continuous and discrete properties to improve the readability of the produced rules. Unlike RxNCM, the CTRENN employs a probabilistic technique to deal with the overlapping of attribute data ranges in various classes. Experiments are carried out using six real life datasets obtained from the UCI repository in order to illustrate the efficacy of the proposed CTRENN algorithm in comparison to the current methods.