Article Preview
TopIntroduction
Even with the advances in the area, students attending e-learning degree programs still drop out at substantially higher rates than their counterparts in on-campus programs (Doulik et al., 2017). This phenomenon has motivated the development of several approaches that intend to predict and understand the reasons for this problem (Pereira & Zambrano, 2017; Choi & Park, 2018; Lykourentzou et al., 2009). These approaches usually first specify a model, most of the time based on data mining techniques (Mouri et al., 2019; Angeli et al., 2017), so that this model can be able to classify if a student tends to drop out. However, current proposals fail in considering the temporal aspect of the majority of the attributes. Rather, only specific moments are considered and predictions that are calculated in such moments tend to lose their accuracy when analyzed over a timeline.
The limitation in predicting the dropout intention of a student, based only on a timeline moment, is straightforward. At a given moment, a set of attributes have a higher influence on the final result. When the student moves along the terms, such influences are modified. Thus, a prediction taken at the beginning will not have the same accuracy at a later moment if the same model is employed. Thus, a prediction system should adjust its model considering the moment that such a prediction is taken. Rule-based classification is a proper technique to be applied to this problem due to two features. First, the generation of rules directly highlights the attributes that are more important at each moment. Thus, the process just needs to select such attributes rather than evolving or adjusting previous attributes according to the new context. Second, both attributes and relations between attributes are clearly specified since the reasons of decisions are represented by means of rules, different from Neural Network or Support Vector Machine approaches (Flanagan & Hirokawa, 2018), which deliver result without an explanation of how the results were obtained.
The definition of any prediction system must include its prediction model. As this work intends to customize models to different moments, then a first step is to identify the attributes that are essential to each of these moments. Important attributes in this context are related to the student performance, as demonstrated in previous works (Pal & Pal, 2013). However, these attributes are only available at the end of each term so that a potential solution is to predict this performance and define strategies to integrate the results of such prediction to the drop out analysis. In this present work, such integration was investigated on the perspective of the rule-based classification techniques because the rationale of decisions can be delivered to the educational staff, better supporting their work. As each technique treats the quality of the generated rules and their accuracy in different ways, the characterization of these techniques, considering the present application domain, was also part of this work.