Student dropout is a serious issue in that it not only affects the individual students who drop out but also has negative impacts on the former university, family, and society together. To resolve this, various attempts have been made to predict student dropout using machine learning. This paper presents a model to predict student dropout at Sahmyook University using machine learning. Academic records collected from 20,050 students of the university were analyzed and used for learning. Various machine learning algorithms were used to implement the model, including Logistic Regression, Decision Tree, Random Forest, Support Vector Machine, Deep Neural Network, and LightGBM (Light Gradient Boosting Machine), and their performances were compared through experiments. We also discuss the influence of oversampling used to resolve data imbalance issues in the dropout data. For this purpose, various oversampling algorithms such as SMOTE, ADASYN, and Borderline-SMOTE were tested. Our experimental results showed that the proposed model implemented using LightGBM provided the best performance with an F1-score of 0.840, which is higher than the results of previous studies discussing the dropout prediction with the issue of class imbalance.

The F1 score has been widely used to measure the performance of machine learning models. However, it is variant to the ratio of the positive class in the training data, π. Depending on how large π is, it can be underestimated or overestimated, making it difficult to fairly compare the performance of models. In this study, we discuss how to calibrate the F1 score for fair performance comparison of binary classification models trained on data with different positive class ratios. We initially demonstrated that the F1 score is inverse proportional to accuracy according to the change in π. From the relationship, the calibrated F1 score was defined as an arithmetic mean of the two measures, which we named the F1* score. Since many prior studies only presented the F1 score or accuracy for model performance, but not both, we provided additional equations to estimate the expected F1 score or accuracy when one of the two measures is available. The accuracy of the presented equations was examined through experiments with a real dataset aimed at student dropout prediction, and the results showed that the mean absolute difference between the derived and actual values was less than 0.01, inferring that the proposed F1* score can calibrate a given F1 score with a high level of accuracy. We also conducted an example analysis comparing the performance of existing models using the F1* score to highlight its efficacy.

User reviews have been widely used to extract opinions, complaints, and requirements about a given place or product from the users’ point of view. In the process of collecting user reviews, a lot of reviews irrelevant to a given search keyword are included in the collection result. Such irrelevant reviews can easily be detected using supervised learning algorithms. However, situation changes when the number of places or products that need to be analyzed increases because manual labeling is required for each collected review. This paper presents a method to detect irrelevant reviews efficiently when a large number of places and reviews need to be analyzed. The basic idea of the proposed method is to expand the target of the learning from an individual place to a group consisting of multiple places. The method can be applied properly to any place whose number of reviews is not enough to perform the training because a classifier obtained by training sufficient reviews included in the places of a group can be used for the places with insufficient reviews in the group. As an initial study of this approach, we tried to check through experiments whether the proposed method can provide higher accuracy than the conventional method where the training is performed for individual places. Our experimental results showed that the average f1-score of the group learning was about 0.931 over real data collected online. For the places with less than 100 reviews, the f1-score of the conventional and proposed methods was 0.651 and 0.865, respectively, showing that 21.4% performance improvement.