Scholarlink

In the domain of psychopathology, the use of Ecological Momentary Assessment (EMA) allows for real-time data collection, providing valuable insights into the temporal dynamics and variability across multiple individuals. Such rich EMA information can be used to build personalized models that predict the future of EMA symptomatology. However, a limited amount of data points is usually available for each individual. To address this, transfer learning approaches can be applied to improve predictions for a specific individual (target domain) by incorporating data from other individuals (source domain). Among the existing transfer learning approaches, boosting-based methodologies, focusing on Transfer Adaptive Boosting (TrAdaBoost), are further explored. Specifically, after identifying the issues of the original TrAdaBoost, we propose various enhancements in the modeling process. To evaluate the effectiveness of all the proposed enhancements, such as the optimal selection of similar source domains and their weighting strategies, their impact on performance is investigated. The findings highlight that an initial source weight increase is a necessary step, while integrating more sources gives better results. Subsequently, these are compared against several baselines, including personalized models and the original TrAdaBoost, showing an improved performance compared to the latter case. Thus, our work contributes to the development of a more effective transfer learning framework for EMA data, ultimately improving the prediction of mental health symptomatology.

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

Enhanced Boosting-Based Transfer Learning for Modeling Ecological Momentary Assessment Data

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Mandani Ntekouli,
  • Gerasimos Spanakis,
  • Lourens Waldorp,
  • Anne Roefs

摘要

In the domain of psychopathology, the use of Ecological Momentary Assessment (EMA) allows for real-time data collection, providing valuable insights into the temporal dynamics and variability across multiple individuals. Such rich EMA information can be used to build personalized models that predict the future of EMA symptomatology. However, a limited amount of data points is usually available for each individual. To address this, transfer learning approaches can be applied to improve predictions for a specific individual (target domain) by incorporating data from other individuals (source domain). Among the existing transfer learning approaches, boosting-based methodologies, focusing on Transfer Adaptive Boosting (TrAdaBoost), are further explored. Specifically, after identifying the issues of the original TrAdaBoost, we propose various enhancements in the modeling process. To evaluate the effectiveness of all the proposed enhancements, such as the optimal selection of similar source domains and their weighting strategies, their impact on performance is investigated. The findings highlight that an initial source weight increase is a necessary step, while integrating more sources gives better results. Subsequently, these are compared against several baselines, including personalized models and the original TrAdaBoost, showing an improved performance compared to the latter case. Thus, our work contributes to the development of a more effective transfer learning framework for EMA data, ultimately improving the prediction of mental health symptomatology.