Quantum inspired genetic algorithm with clustering-based bagging approach for improved deceit classification

Abstract

Deception detection is a process of identifying a person's truthfulness and credibility. It is often used in criminal analysis. Ensemble-based models have shown potential in this field, but they struggle with selecting diversity-based classifiers to improve performance. Previous attempts failed to reduce the ensemble model complexity without compromising model performance. To address this, a novel strategy that blends a Quantum-Inspired Genetic Algorithm (QGA) with Integrated K-means clustering is proposed. The QGA leverages the quantum operators for better exploration and exploitation. The clustering-based approach enhances the diversity in data for better generalization capability. In our model, integrated clustering is applied for sample grouping. These samples are used to generate bootstrap bags. Every bag is applied on four diverse learners, namely Decision Tree, Support Vector Machine, Multi-Layer Neural Network, and Naïve Bayes to generate 40 learners. Further, QGA with novel fitness function employed to select the best classifiers from the pool. To avoid premature convergence and maintain diversity in QGA-based approach, experiments are run twenty times with different random initial populations. Our proposed model is tested on our own dataset collected using the Concealed Information Test. The proposed model outperformed when compared with baseline methods, State-Of-The-Art ensemble and deep learning methods. To prove the robustness of the proposed model, an ablation and leave-one-subject vs others experiment were conducted. From these experiments, it's been clear that the hybridization of Cluster Validity Index, clustering, bootstrapping, and QGA not only increased performance and generalization capability but also reduced ensemble complexity.