Abstract
Academic Performance Prediction models mustn't be accurate only, but timely too, to identify at-risk students at the earliest to provide remedy. Heart rate data of 50 students in 3 main courses are collected, processed, and analyzed to distinguish the difference between excellent students and at-risk students. Three of the 12 heart rate attributes were chosen to calculate the threshold values, which are used to predict at-risk students. Half of the at-risk students were identified after week 5. Later, the datasets were rebalanced. Using four Data Mining classifiers, six attributes were identified to be the best attributes for prediction model development. The datasets were then dimensionally reduced. Applying classification, half of the at-risk students were identified earliest around week 5 of the 12-week semester. J48 is the most robust classifier, compared to JRip, Multi-Level-Perceptron, and RandomForest, making accurate prediction on at-risk students earlier most of the time.
TopIntroduction
Educational Data Mining (EDM) is a branch of Data Mining, made popular for more than a decade. Of the many research directions within EDM, predicting students’ performance is a crucial area to arrest the widespread issue of student failure in examination. Student failure and subsequent dropout presents a social dilemma that is growing stealthily. Socially, students who failed or dropped out tend to carry unnecessary psychological burden that is overwhelmingly affecting his or her kin and circles of friends. Today, the teaching fraternity is multitasking in academic and clerical works, leaving them with little or no time to personally care for individual students. There exist many Students’ Performance Prediction (SPP) models that are accurate in identifying at-risk students, especially after incorporating academic data such as internal assessment results. However, the teachers are left with an impossible task of providing remedial help to such students as little time is remaining before the final examination.
One suggestion to identify at-risk students faster is to use head-mounted MRI (Magnetic Resonance Imaging) scanner to detect whether a student is cognitively stimulated during class hours to determine whether he or she is engaged in learning, thus gauging the feasibility of passing the subject. However, this suggestion proves to be expensive and cumbersome, as the device is sophisticated and wired. With the proliferation of wearable wireless physiological health trackers, it presents a new dimension to evaluate level of class engagement of students, with a potential to predict the final academic performance of students. However, the current technological advancement in health trackers only enables heart rate to be collected accurately. Other parameters such as blood pressure and oxygen saturation are yet to be accurately measured. This scenario intrigued the authors to commence the research to use heart rate data to measure the engagement level of students during class and then develop a model to predict student performance.
The initial plan was to determine if there is a difference in the heart rate pattern between academically high performing students and low performing students. A high performing student is one who can pass his final exam while a low performing student is one who will fail in his final exam. This is done by observing the number of Spike from graphs plotted using the heart rate data. Heart rate fluctuation can be in the form of a sudden surge or a sudden drop of heart rate. If there is a sudden surge and then followed by a sudden drop of heart rate, or vice versa, then it forms a spike upward or downward, respectively. Then, the heart rate patterns are analyzed to determine threshold values of heart rate fluctuation which can be used to predict the future performance of students in their examinations. Applying the threshold values in the prediction model to test the datasets arranged in cumulative weekly manner, the aim is to determine how timely will the model be accurate. A comparison of various heart rate parameters (such as maximum heart rate, minimum heart rate, average heart rate, etc.) is then analyzed using Leave-One-Out method to determine the most important attributes in the model prediction development. Using the best set of attributes to develop a SPP model based on four classification algorithms, and then testing them with the cumulative weekly datasets, the authors will obtain the second set of timeliness results. Ultimately, we shall decide whether the SPP model developed using threshold values or the SPP model developed using best set of attributes is better in terms of timeliness. In short, the objectives of this study can be summarized as below:
- 1.
Determine that the heart rate pattern of high performing students and low performing students are different based on number of Spike.
- 2.
Determine the threshold values of Surge, Drop and Spike that can be used to develop a SPP model to predict whether a student will pass or fail in the final examination.
- 3.
Determine the most important attributes from the various heart rate attributes that should be used to develop a SPP model.
- 4.
Compare the timeliness factor of the SPP model developed using threshold values and the SPP model developed using best set of attributes, to decide which is better.
Key Terms in this Chapter
Class Rebalancing: A method to increase a minority class or to reduce a majority class in order that classes have the same number of instances.
Heart Rate Fluctuation: The pattern of heart rate increasing or decreasing over time.
Dimension Reduction: The process of removing redundant or dependent attribute(s) in order to increase prediction model accuracy and to reduce model development time.
Accuracy: The percentage of both passed students and failed students combined predicted accurately.
Classifier: A supervised Data Mining algorithm used to categorize an instance into one of the two or more classes.
Recall: The percentage of all failed students that are predicted to fail.
At-risk Students: Students who are failing or about to fail.
Remedial Help: Extra tutorial help given to academically weaker students.
Ensemble-based: Combining two or more algorithms into one.