Article Preview
TopIntroduction
Machine learning techniques have been extensively utilized in various areas of education and instructional practices. For instance, boosting algorithm has been employed to predict student performance in mathematics(Hamim, et al., 2022). Additionally, deep learning models have been utilized to analyze and model big data pertaining to college student employment management(Yin, 2023). Furthermore, machine learning technologies have been applied in the field of classical music education(Wang, 2023).
In the specific realm of modeling students' performance, feature engineering plays a crucial role in enhancing the accuracy of predictive models. Feature engineering involves selecting, extracting, and transforming relevant features from the data to improve the performance of machine learning algorithms(Gupta & Gueneau, 2021). By deriving features that represent interactions between different variables, such as parent relationship status and travel time between home and school, feature engineering can positively impact student academic performance predictions(Gupta & Gueneau, 2021).
Moreover, the fusion of multiple features, such as student behavior features and exercise features, using attention mechanisms has been proposed as a novel framework for student performance prediction(Liu et al., 2020). This approach aims to leverage a combination of different types of features to enhance the accuracy of predictive models.
Furthermore, feature selection methods have been highlighted as critical in developing accurate student performance prediction models(Zaffar et al., 2022). Utilizing techniques like FCBF feature selection can aid in identifying the most relevant features for predicting student outcomes(Zaffar et al., 2020).
In conclusion, feature engineering, encompassing the selection, extraction, and transformation of relevant features, is a fundamental aspect of predicting students' performance accurately. By incorporating various types of features, including demographic, academic, behavioral, and exercise-related features, predictive models can be enhanced to provide valuable insights into student outcomes.
The majority of extant research has predominantly focused on the domain of primitive features—that is, features that are to be selected, extracted, and combined, which are largely uncontroversial and inherent, such as demographic characteristics and family background attributes. However, we posit that in predicting learning achievement, it is imperative to initiate a more profound investigation into the essence of 'learning' itself. Consequently, we have directed our research focus toward the potential enhancement in predictive accuracy that can be attributed to the incorporation of 'learning style' as a predictive feature.
A recent study delves into teaching control theory by selecting methodologies based on various well recognized learning styles, including the Kolb, Felder and Silverman, and VARK schemes(Rojas-Palacio et al., 2022). The plethora of available schemes for selection is precisely indicative of the absence of a unified definition. 'Learning style' is not an innate characteristic; rather, its definition is inherently ambiguous and warrants further exploration.
However, our research does not endeavor to delineate 'learning style', but rather aims to capitalize on the extant data at our disposal, abstracting a feature named 'learning style' as an intermediary step in the modeling process for the prediction of student academic performance—it is a means to an end, not the end itself. In pursuit of elucidating the optimal representation of the 'learning style' attribute, our study employs a dual approach, leveraging both machine learning-based automatic clustering algorithms and manual labeling methodologies.
The objective of our research is to investigate methodologies that enhance the precision of regressors in the task of predicting student academic performance. Specifically, we aim to determine whether the inclusion of 'learning style' can augment the accuracy of predictions. Furthermore, we seek to evaluate the relative efficacy of manual feature engineering versus features labeled with automated clustering.