The purpose of learning analytics is to promote student success in the classroom. To implement the framework of learning analytics, researchers have adopted machine learning methodologies to identify at-risk students at an early stage. In theory, machine learning is a mathematical algorithm that improves automation through experience. The experience is the data collected from online learning platforms, and in general, the data contain various features such as the number of times that a student accesses the learning material each week. Relevant studies have demonstrated extremely high accuracy in identifying at-risk students using identification models trained by machine learning. However, numerous details and data challenges have been overlooked in prior studies, calling into question the accuracy of past contributions. In this study, we focused on one type of data challenge: data imbalance. The data imbalance problems in education are usually the result of teachers’ grading policy. To highlight the seriousness of this issue, we collected data from 12 blended learning courses and summarized 3 types of grading policies: discrimination, stringency, and leniency. We then provided evidence that the leniency strategy causes the illusion of high accuracy of at-risk student identification. Finally, we verified a robust method to address the effectiveness of the leniency strategy, and using these results, we summarized the characteristics of students who tend to be misidentified by machine learning methodology.

学习分析的目的是促进学生在课堂上的成功。为了实施学习分析的框架，研究人员采用了机器学习方法来在早期识别高风险学生。从理论上讲，机器学习是一种数学算法，可以通过经验来提高自动化程度。体验是从在线学习平台收集的数据，通常，数据包含各种功能，例如学生每周访问学习材料的次数。相关研究表明，使用通过机器学习训练的识别模型，识别高危学生的准确性非常高。然而，在先前的研究中，许多细节和数据挑战被忽略了，这使人们对过去贡献的准确性提出了质疑。在这个研究中，我们专注于一种类型的数据挑战：数据不平衡。教育中的数据不平衡问题通常是教师评分政策的结果。为了强调此问题的严重性，我们从12个混合学习课程中收集了数据，并总结了三种评分策略：歧视，严格和宽大处理。然后，我们提供了证据表明宽大处理策略导致了高风险学生识别准确性的幻觉。最后，我们验证了解决宽恕策略有效性的有效方法，并使用这些结果，总结了倾向于被机器学习方法误认的学生的特征。为了强调此问题的严重性，我们从12个混合学习课程中收集了数据，并总结了三种评分策略：歧视，严格和宽大处理。然后，我们提供了证据表明宽大处理策略导致了高风险学生识别准确性的幻觉。最后，我们验证了解决宽恕策略有效性的有效方法，并使用这些结果，总结了倾向于被机器学习方法误认的学生的特征。为了强调此问题的严重性，我们从12个混合学习课程中收集了数据，并总结了三种评分策略：歧视，严格和宽大处理。然后，我们提供了证据表明宽大处理策略导致了高风险学生识别准确性的幻觉。最后，我们验证了解决宽恕策略有效性的有效方法，并使用这些结果，总结了倾向于被机器学习方法误认的学生的特征。