Dynamic-unit discrete hopfield neural network with supervised preprocessing phase: optimizing the logic mining using S-type random 2 satisfiability for classification

Logic mining, a specialized area of data mining, extracts logical rules from datasets by examining their underlying behavior. However, a critical limitation of many existing models is their insufficient focus on ensuring the quality of processed data for knowledge extraction. This often leads to challenges in the transparency and reliability of derived outcomes. Consequently, maintaining robust data quality, reliability, and availability is essential for producing meaningful and interpretable knowledge. This paper introduces a novel logic mining model designed to enhance both the quantity and quality of induced logic by capturing all optimal solutions. The proposed model incorporates a supervised preprocessing phase that identifies optimal attributes using Elastic Net regularization for robust feature selection. To enhance logical diversity, a non-systematic S-Type Random 2 Satisfiability structure is introduced, guided by probability distributions that reflect dataset characteristics. This probabilistic logic operates within a novel Dynamic-Unit Discrete Hopfield Neural Network, which ensures comprehensive solution space coverage through dynamic Content Addressable Memory units, leading to the production of diverse and optimal induced logical rules. Furthermore, a newly developed shifting operator in the retrieval phase facilitates the optimal arrangement of attributes in the final induced logic. Experimental evaluations demonstrate that the proposed model significantly outperforms existing approaches, achieving Accuracy 82 %, Specificity of 99.8 % and a Negative Predictive Value of 85.6 %. These findings highlight the effectiveness of the supervised preprocessing phase, the implementation of a non-systematic probabilistic structure, and the efficiency of the dynamic unit network in advancing logic mining capabilities.

Text 122839.pdf - Published Version Restricted to Repository staff only Download (13MB) Official URL or Download Paper: https://www.sciencedirect.com/science/article/pii/...

Actions (login required)

View Item