Scholarlink

<p>Social media sentiment analysis benefits from deep models, yet personalization remains limited when user-specific behavior and platform heterogeneity are underused. This study presents a novel personalized feature fusion framework that integrates user historical behavior with BERT-based textual analysis, employing three complementary feature fusion mechanisms: an adaptive gated module that dynamically filters irrelevant features, an attention-based fusion component that balances user histories with textual cues, and an efficient concatenation strategy that preserves comprehensive behavioral patterns as a standard baseline. Our key innovation is an explicit user-history fusion mechanism that injects historical behavior signals directly into the sentiment classifier, ensuring consistent personalization across user activity levels. The design significantly reduces the semantic gap between textual features and behavioral patterns through a unified optimization objective, achieving robust cross-domain adaptability. Comprehensive experiments on Sentiment140 and Yelp datasets demonstrate the framework’s superiority, with the gated fusion approach achieving 94.26% accuracy across datasets. Under the standardized Sentiment140 protocol with <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\le 50\)</EquationSource> </InlineEquation> posts per user, the gated fusion model attains 92.42% accuracy, compared with the 86.5–87.1% range reported by related personalization baselines. Ablation studies verify the essential contribution of each component, particularly in handling sparse user data. The framework shows promising potential in product review analysis and social media monitoring applications, while providing practical guidelines for implementing personalized sentiment analysis in real-world scenarios. Despite these gains, the approach still depends on a modest amount of user history and periodic recalibration of personalization thresholds when expression patterns drift. This work advances the field of personalized sentiment analysis by providing an efficient fusion framework, establishing a foundation for future research in adaptive feature integration and sparse data handling.</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

Personalized sentiment analysis via historical behavior integration: a multi-strategy feature fusion approach

  • 期刊论文

  • 发表日期：

    2026年3月30日

查看全文
  • Kai Yan,
  • Jiaxi Lai

摘要

Social media sentiment analysis benefits from deep models, yet personalization remains limited when user-specific behavior and platform heterogeneity are underused. This study presents a novel personalized feature fusion framework that integrates user historical behavior with BERT-based textual analysis, employing three complementary feature fusion mechanisms: an adaptive gated module that dynamically filters irrelevant features, an attention-based fusion component that balances user histories with textual cues, and an efficient concatenation strategy that preserves comprehensive behavioral patterns as a standard baseline. Our key innovation is an explicit user-history fusion mechanism that injects historical behavior signals directly into the sentiment classifier, ensuring consistent personalization across user activity levels. The design significantly reduces the semantic gap between textual features and behavioral patterns through a unified optimization objective, achieving robust cross-domain adaptability. Comprehensive experiments on Sentiment140 and Yelp datasets demonstrate the framework’s superiority, with the gated fusion approach achieving 94.26% accuracy across datasets. Under the standardized Sentiment140 protocol with \(\le 50\) posts per user, the gated fusion model attains 92.42% accuracy, compared with the 86.5–87.1% range reported by related personalization baselines. Ablation studies verify the essential contribution of each component, particularly in handling sparse user data. The framework shows promising potential in product review analysis and social media monitoring applications, while providing practical guidelines for implementing personalized sentiment analysis in real-world scenarios. Despite these gains, the approach still depends on a modest amount of user history and periodic recalibration of personalization thresholds when expression patterns drift. This work advances the field of personalized sentiment analysis by providing an efficient fusion framework, establishing a foundation for future research in adaptive feature integration and sparse data handling.