Abstract:

Massive Open Online Courses (MOOCs) have gained rising popularity in recent years. However, MOOCs have faced a challenge of a large number of students dropping out from courses. Most studies predict dropouts based on some general features extracted from historical learning behavior and ignore the diversity of the behaviors. To solve this problem, we first analyze each type of learning behavior independently to get the different behavior patterns between dropout and retention students. We then derive multiple kinds of features from the corresponding types of learning behavior records. After that, we propose three algorithms that make use of these features. The first one trains several detectors based on each types of features. The second utilizes multi-view ensemble learning to anticipate dropouts. The third applies semi-supervised co-training to train the detector. Experimental results justify the rationality of the multi-view features and the proposed approaches achieve better prediction performances.