airo 23(1):

Research Article

Multi-class Classification of Imbalanced Intelligent Data using Deep Neural Network

Download170 downloads
  • @ARTICLE{10.4108/airo.3486,
        author={Masoumeh Soleimani and Akram Sadat  Mirshahzadeh},
        title={Multi-class Classification of Imbalanced Intelligent Data using Deep Neural Network},
        journal={EAI Endorsed Transactions on AI and Robotics},
        volume={2},
        number={1},
        publisher={EAI},
        journal_a={AIRO},
        year={2023},
        month={7},
        keywords={Imbalanced Data, Dynamic Sampling, Deep Neural Network},
        doi={10.4108/airo.3486}
    }
    
  • Masoumeh Soleimani
    Akram Sadat Mirshahzadeh
    Year: 2023
    Multi-class Classification of Imbalanced Intelligent Data using Deep Neural Network
    AIRO
    EAI
    DOI: 10.4108/airo.3486
Masoumeh Soleimani1,*, Akram Sadat Mirshahzadeh2
  • 1: Clemson University
  • 2: Islamic Azad University, Isfahan
*Contact email: m.soleimani90@gmail.com

Abstract

In recent years, studies in the field of deep learning have made significant progress. These studies have focused more on datasets with balanced classification, and less research has been done on imbalanced datasets, which are of great importance in the real world and present significant challenges for classification. This article studies the problem of classifying imbalanced data, introduces dynamic sampling for deep neural networks, investigates the imbalanced multiclass problem, and proposes a dynamic sampling method for deep learning. In our proposed method, all samples are fed to the current deep neural network for each training iteration, and the accuracy, precision, and mean error of the deep neural network are estimated. The proposed method dynamically selects informative data for training the deep neural network. Comprehensive experiments were conducted to evaluate and understand its strengths and weaknesses. The results of 13 imbalanced multiclass datasets show that the proposed method outperforms other methods, such as initial sampling techniques, active learning, cost-sensitive learning, and reinforcement learning.