Scholarlink

<p>This study presents an optimized recurrent neural network architecture, termed “RegGRU-Opt”, designed for binary classification tasks such as sentiment analysis on Amazon reviews. By integrating gated recurrent units (GRUs), strategic dropout layers, L2 regularization in dense layers, and advanced optimization techniques (e.g., early stopping and learning rate scheduling), the proposed model achieves robust generalization and outperforms baseline models like CNN, RNN, and SeqClassRNN. Experimental results show high accuracy (≈98%), F1-Score (≈97%), and ROC AUC (≈99%), underscoring the model’s capacity to capture nuanced textual patterns. Data balancing methods, including SMOTE and SMOTE + Tomek Links, ensure fair representation of both classes, while feature importance analyses (Chi-squared and SHAP) enhance model interpretability. Despite improved performance, challenges remain in terms of domain adaptation, computational resources, and handling imbalanced data. Future work will focus on exploring hybrid models, advanced regularization methods, broader datasets, and improved efficiency, ultimately expanding the applicability of RegGRU-Opt across diverse real-world text classification scenarios.</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

RegGRU-Opt: A Robust GRU-Based RNN Model for High-Performance Sentiment Analysis and Binary Classification

  • 期刊论文

  • 发表日期：

    2025年1月31日

查看全文
  • Aashita Chhabra,
  • Mansaf Alam

摘要

This study presents an optimized recurrent neural network architecture, termed “RegGRU-Opt”, designed for binary classification tasks such as sentiment analysis on Amazon reviews. By integrating gated recurrent units (GRUs), strategic dropout layers, L2 regularization in dense layers, and advanced optimization techniques (e.g., early stopping and learning rate scheduling), the proposed model achieves robust generalization and outperforms baseline models like CNN, RNN, and SeqClassRNN. Experimental results show high accuracy (≈98%), F1-Score (≈97%), and ROC AUC (≈99%), underscoring the model’s capacity to capture nuanced textual patterns. Data balancing methods, including SMOTE and SMOTE + Tomek Links, ensure fair representation of both classes, while feature importance analyses (Chi-squared and SHAP) enhance model interpretability. Despite improved performance, challenges remain in terms of domain adaptation, computational resources, and handling imbalanced data. Future work will focus on exploring hybrid models, advanced regularization methods, broader datasets, and improved efficiency, ultimately expanding the applicability of RegGRU-Opt across diverse real-world text classification scenarios.