Scholarlink

<p>The detection of drifting sentiment in public opinion remains a complex and underexplored problem, largely due to three fundamental challenges: the high cost of manual annotation, the inherent imbalance of sentiment distributions, and the temporal dynamics that drive shifts in public discourse. Traditional supervised learning approaches struggle with scalability and robustness in such environments, making them less effective for real-world applications. To address these issues, this paper proposes a novel weakly supervised sentiment drift detection framework that integrates contextual embeddings, data rebalancing, and ensemble classification. Specifically, BERT embeddings are employed to capture deep semantic features, TF-IDF is incorporated to preserve complementary statistical information, SMOTE is applied to alleviate class imbalance by generating synthetic minority samples, and XGBoost is utilized as a robust and efficient classifier that performs well under noisy and weakly labeled conditions. Extensive experiments conducted on the ThaiCaveRescue and IMDB datasets demonstrate that the proposed BERT–SMOTE–XGBoost framework consistently outperforms baseline and state-of-the-art methods in terms of accuracy, balanced accuracy, and weighted F1 score, while maintaining competitive computational efficiency. Ablation studies further validate the contribution of each component, showing that the removal of TF-IDF, weak supervision, SMOTE, or XGBoost leads to notable performance degradation. Moreover, hyperparameter sensitivity analysis on feature sampling rate, learning rate, and the number of decision trees confirms the robustness and stability of the proposed approach across different configurations. Overall, the results highlight that combining deep semantic representations with classical statistical features, rebalancing techniques, and ensemble-based classification provides a practical and scalable solution for sentiment drift detection. This work not only advances the methodological foundation of sentiment analysis under weak supervision but also offers practical insights for deploying robust opinion monitoring systems in dynamic, real-world contexts.</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

A weakly supervised SMOTE-XGBoost framework with sliding-window monitoring for temporal sentiment drift detection

  • 期刊论文

  • 发表日期：

    2026年6月16日

查看全文
  • Chuanjun Zhao,
  • Xiaoxiong Xi,
  • Lu Kang,
  • Junqiang Bai

摘要

The detection of drifting sentiment in public opinion remains a complex and underexplored problem, largely due to three fundamental challenges: the high cost of manual annotation, the inherent imbalance of sentiment distributions, and the temporal dynamics that drive shifts in public discourse. Traditional supervised learning approaches struggle with scalability and robustness in such environments, making them less effective for real-world applications. To address these issues, this paper proposes a novel weakly supervised sentiment drift detection framework that integrates contextual embeddings, data rebalancing, and ensemble classification. Specifically, BERT embeddings are employed to capture deep semantic features, TF-IDF is incorporated to preserve complementary statistical information, SMOTE is applied to alleviate class imbalance by generating synthetic minority samples, and XGBoost is utilized as a robust and efficient classifier that performs well under noisy and weakly labeled conditions. Extensive experiments conducted on the ThaiCaveRescue and IMDB datasets demonstrate that the proposed BERT–SMOTE–XGBoost framework consistently outperforms baseline and state-of-the-art methods in terms of accuracy, balanced accuracy, and weighted F1 score, while maintaining competitive computational efficiency. Ablation studies further validate the contribution of each component, showing that the removal of TF-IDF, weak supervision, SMOTE, or XGBoost leads to notable performance degradation. Moreover, hyperparameter sensitivity analysis on feature sampling rate, learning rate, and the number of decision trees confirms the robustness and stability of the proposed approach across different configurations. Overall, the results highlight that combining deep semantic representations with classical statistical features, rebalancing techniques, and ensemble-based classification provides a practical and scalable solution for sentiment drift detection. This work not only advances the methodological foundation of sentiment analysis under weak supervision but also offers practical insights for deploying robust opinion monitoring systems in dynamic, real-world contexts.