Scholarlink

<p>In the domain of learning analytics, which leverages data-driven approaches to support teaching, a valuable application emerges in the form of AI-based educational predictive models. Beyond performance, researchers increasingly emphasize the fairness of such models, with fairness being assessed based on demographic attributes. This paper first constructs an engagement detection model within an online learning community exhibiting gender imbalances, situated in a blended learning university environment psychology course. Drawing upon communication theory, the model’s features are designed, followed by an exploration of statistical disparities between male and female characteristics. Subsequently, through a comparative analysis of AI models that either exclude or include gender, considering performance changes and group fairness, coupled with explainable AI methodologies, the impact of gender on the models is examined. The results suggest that improvements in model performance are more likely to benefit from sensitive attributes with high feature importance. In scenarios where gender is excluded, latent biases inherent within the dataset may lead to group unfairness, exacerbated upon inclusion of gender. Based on these findings, we delineate circumstances under which modeling sensitive attributes is permissible, while discussing arguments supporting and opposing the inclusion of such attributes from both performance and fairness perspectives. The analytical methods and findings of this study offer novel insights for designing more equitable and effective educational predictive models within learning analytics.</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

Should educational AI models include gender attribute? explaining the why based on environmental psychology course with gender imbalance

  • 期刊论文

  • 发表日期：

    2025年9月12日

查看全文
  • Hai Li,
  • Wanli Xing,
  • Wangda Zhu,
  • Shan Zhang,
  • Zifeng Liu

摘要

In the domain of learning analytics, which leverages data-driven approaches to support teaching, a valuable application emerges in the form of AI-based educational predictive models. Beyond performance, researchers increasingly emphasize the fairness of such models, with fairness being assessed based on demographic attributes. This paper first constructs an engagement detection model within an online learning community exhibiting gender imbalances, situated in a blended learning university environment psychology course. Drawing upon communication theory, the model’s features are designed, followed by an exploration of statistical disparities between male and female characteristics. Subsequently, through a comparative analysis of AI models that either exclude or include gender, considering performance changes and group fairness, coupled with explainable AI methodologies, the impact of gender on the models is examined. The results suggest that improvements in model performance are more likely to benefit from sensitive attributes with high feature importance. In scenarios where gender is excluded, latent biases inherent within the dataset may lead to group unfairness, exacerbated upon inclusion of gender. Based on these findings, we delineate circumstances under which modeling sensitive attributes is permissible, while discussing arguments supporting and opposing the inclusion of such attributes from both performance and fairness perspectives. The analytical methods and findings of this study offer novel insights for designing more equitable and effective educational predictive models within learning analytics.