Scholarlink

<p>Multi-layer perceptron (MLP) is a significant technique of artificial neural networks (ANN), which consists of different numbers of hidden layers and neurons. The network can capture intricate patterns and hierarchical features with more hidden layers. Practically, an MLP’s learning ability and performance are greatly influenced by the number of hidden layers and neurons per layer. Thus, selecting an optimal architecture for MLP networks remains a challenge. In these networks, as in machine learning models, the selection feature process is essential to enhance the performance of those models. While entropy is a well-known method as a measure of systemic uncertainty used in information theory. This paper proposes a novel structure to optimize the performance of the MLP model to enhance the classification accuracy. Firstly, a feature selection method is proposed based on the normalized entropy that computes the feature’s degree of diversity in relation to its maximal potential. Then, a dynamic MLP structure is used through three hidden layers with neuron numbers [1-151] for each hidden layer. The best configuration that produces the highest accuracy is chosen. Furthermore, another two classic MLP models are used in this paper. The first one is a classic MLP, while the second is a classic MLP that applied the recursive feature elimination (RFE) as a feature selection method. Five datasets are used to evaluate this model: Heart disease, PIMA, Wine quality, Liver disorders and finally, soybean datasets. The model’s performance is evaluated by different measurements, which are accuracy, precision, recall, F1-score, specificity, Matthews Correlation Coefficient (MCC), and finally, Area Under the Curve (AUC). The findings prove the effectiveness of the proposed MLP model that achieves an accuracy of 93.33%, 81.82%, 82.5%, 71.01%, and 100.0% for these datasets, respectively, outperforming the classic MLP model and many other previous models.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhancing MLP Classification Performance Through Normalized Entropy-based Feature Selection and Dynamic Architecture

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月27日

查看全文
  • Nor Samsiah Sani,
  • Mudatheer M. Al-Slivani,
  • Mayameen S. Kadhim,
  • Ahmed Dheyaa Radhi,
  • Hussein A. A. Al-Khamees,
  • Zulaiha Ali Othman,
  • Mohd Aliff Afira Sani,
  • Rusul Mansoor Al-Amri,
  • Mohammed Amin Almaiah

摘要

Multi-layer perceptron (MLP) is a significant technique of artificial neural networks (ANN), which consists of different numbers of hidden layers and neurons. The network can capture intricate patterns and hierarchical features with more hidden layers. Practically, an MLP’s learning ability and performance are greatly influenced by the number of hidden layers and neurons per layer. Thus, selecting an optimal architecture for MLP networks remains a challenge. In these networks, as in machine learning models, the selection feature process is essential to enhance the performance of those models. While entropy is a well-known method as a measure of systemic uncertainty used in information theory. This paper proposes a novel structure to optimize the performance of the MLP model to enhance the classification accuracy. Firstly, a feature selection method is proposed based on the normalized entropy that computes the feature’s degree of diversity in relation to its maximal potential. Then, a dynamic MLP structure is used through three hidden layers with neuron numbers [1-151] for each hidden layer. The best configuration that produces the highest accuracy is chosen. Furthermore, another two classic MLP models are used in this paper. The first one is a classic MLP, while the second is a classic MLP that applied the recursive feature elimination (RFE) as a feature selection method. Five datasets are used to evaluate this model: Heart disease, PIMA, Wine quality, Liver disorders and finally, soybean datasets. The model’s performance is evaluated by different measurements, which are accuracy, precision, recall, F1-score, specificity, Matthews Correlation Coefficient (MCC), and finally, Area Under the Curve (AUC). The findings prove the effectiveness of the proposed MLP model that achieves an accuracy of 93.33%, 81.82%, 82.5%, 71.01%, and 100.0% for these datasets, respectively, outperforming the classic MLP model and many other previous models.