Scholarlink

Traditional movie recommendation systems often overlook the critical role of user emotion, creating a personalization gap that impacts user satisfaction. This study aimed to address this limitation by developing and evaluating a context-sensitive emotion-based movie recommendation system capable of integrating emotional intelligence. Using the TMDB 5000 dataset, a fine-tuned transformer-based deep learning model was employed to extract seven-dimensional emotion vectors from textual movie overviews, creating a framework to generate recommendations that could either match or regulate a user’s specified emotional state. The results demonstrated a strong performance in emotional alignment (mean Precision@5 of 0.73) and content diversity (0.57). However, the evaluation also revealed significant challenges, including severe class imbalance within the dataset and very low model stability, indicating a high sensitivity to data variations. In conclusion, the findings confirm that affective modeling has significant potential to improve recommendations, but its practical reliability is currently hindered by data-inherent issues, highlighting the critical need to address the imbalance of data sets and the robustness of the model in future work.

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

Emotion-Aware Movie Recommendation Using a Transformer-Based Deep Learning Approach

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Soumashree Das,
  • Jhanvi Murarka,
  • Ishita Dutta,
  • Jhalak Dutta

摘要

Traditional movie recommendation systems often overlook the critical role of user emotion, creating a personalization gap that impacts user satisfaction. This study aimed to address this limitation by developing and evaluating a context-sensitive emotion-based movie recommendation system capable of integrating emotional intelligence. Using the TMDB 5000 dataset, a fine-tuned transformer-based deep learning model was employed to extract seven-dimensional emotion vectors from textual movie overviews, creating a framework to generate recommendations that could either match or regulate a user’s specified emotional state. The results demonstrated a strong performance in emotional alignment (mean Precision@5 of 0.73) and content diversity (0.57). However, the evaluation also revealed significant challenges, including severe class imbalance within the dataset and very low model stability, indicating a high sensitivity to data variations. In conclusion, the findings confirm that affective modeling has significant potential to improve recommendations, but its practical reliability is currently hindered by data-inherent issues, highlighting the critical need to address the imbalance of data sets and the robustness of the model in future work.