Forecasting Using Rules Extracted from Privacy Preservation Neural Network

Abstract

Privacy preserving data mining is of paramount importance in many areas. In this paper, we employ Particle Swarm Optimization (PSO) trained Auto Associative Neural Network (PSOAANN) for preservation privacy in input feature values. The privacy preserved input features are fed to the Dynamic Evolving Neuro Fuzzy Inference System (DENFIS) and Classification and Regression Tree (CART) separately for rule extraction purpose. We also propose a new feature selection method using PSOAANN. Thus, in this study, PSOAANN accomplishes privacy preservation as well as feature selection. The performance of the hybrid is tested using 10 fold cross validation on 5 regression datasets viz. Auto MPG , Body Fat , Boston Housing , Forest Fires and Pollution . The study demonstrates the effectiveness of the proposed approach in generating accurate regression rules with and without feature selection. The ttest at 1% level of significance is performed to see whether the difference in results obtained in the case of with and without feature selection is statistically significant or not. In the case of PSOAANN + CART, it is observed that the result is statistical insignificant between with and without feature selection in four datasets. In the case of PSOAANN + DENFIS, it is observed that statistical significance between with and without feature selection for three datasets. Hence, from the t-test it is concluded that the proposed feature selection method yielded better or comparable results.