Nowadays,
it is crucial to identify students at risk of failing courses and/or
dropping out as early as possible. By adopting a large-scale online
testing system [1] such as learning comprehension testing (LCT)
for every class to measure student comprehension of lectures, we
can accumulate information for learning analytics. This paper is
aimed at producing effective learning strategies for students at
risk by utilizing the learning analytics obtained from the online
testing system.
To do this, we have proposed a newly developed method to identify
students likely fail in end-term examination by analyzing the similarity
of the trends of estimated studentsf abilities obtained by item response
theory. The nearest neighbor methodology related to the method of
NNRMLR ([2], e.g.) is adopted for analizing the similarity of learning
skill in the trends of estimated abilities. In determining the optimal
threshold value for success or failure prediction, we have applied
receiver operating characteristic curve and recall precision curve
in precisely investigating the accuracy of the proposed method. Using
data accumulated in 2017 fiscal year, we have constructed a model
and its parameters for a fundamental mathematics subject, resulting
in higher prediction accuracy for 2018 fiscal year success/failure
than that using only the item response theory result with a full
response matrix of M(n,m), where the number of students n is more
than 1,000, and the number of items is almost 100.
Therefore, pedagogical implications focusing on the findings of this
research are the fol- lowing: 1) it is important to accumulate learning
data such as LCT week by week in order to support students at risk,
2) even if only the first half of the ability trends are available,
we can predict the risk of failure at the end-term examination using
the similarities of ability trends more accurately than using only
the LCT abilities via the full response matrix. |
|
|
|
|
success/failure
prediction, large-scale online testing, learning analytics. |
|
|