Analysis and Prediction of At-Risk Students Using Machine Learning Algorithms

Imagine a university as a large, busy garden. The goal is to help every student (the plants) grow strong and reach full bloom (graduation). However, sometimes plants start to wilt or stop growing before they are ready. In the past, gardeners (university staff) would only notice a plant was in trouble after it had already turned yellow or dropped its leaves. By then, it was often too late to save it.

This paper is about building a smart weather forecast system for that garden. Instead of waiting for the leaves to turn yellow, the researchers used computer programs (Machine Learning) to look at the soil, the water, and the history of the plant to predict beforehand which ones might struggle.

Here is a simple breakdown of what they did and what they found:

The Problem: The "Late Alarm"

Universities lose money and reputation when students leave early. Currently, schools try to help students by offering counseling or tutoring, but they often do this reactively. It's like calling a doctor only after you've already broken your leg. The researchers wanted to find a way to predict the break before it happens so they could offer a cast (support) early on.

The Tool: The "Crystal Ball" Algorithms

The researchers gathered data from Sydney International School of Technology and Commerce (SISTC). They looked at 2,405 student records but cleaned them up to focus on 1,027 students who had either finished their studies or dropped out.

They fed this data into four different "crystal balls" (Machine Learning algorithms) to see which one could predict the future best:

1. Logistic Regression: A simple, straight-line calculator that weighs different factors.
2. Random Forest: A team of many decision-makers (like a committee of experts) voting on the outcome.
3. K-Nearest Neighbors (KNN): A method that looks at a student's "neighbors" (students with similar grades and backgrounds) to guess what will happen to them.
4. Support Vector Machines (SVM): A method that draws a line to separate "graduates" from "dropouts."

The Ingredients: What Did They Look At?

To make their predictions, the computer looked at 14 specific clues, such as:

• Academic History: How many subjects they passed, failed, or skipped.
• Demographics: Their age, gender, nationality, and visa type.
• Time: How long they had been studying and how many times they tried a course.

The Results: Who Won the Race?

The researchers tested these "crystal balls" to see how accurate they were. The results were surprisingly high:

• The Winners: Logistic Regression and SVM (Linear Kernel) were the champions. They got 99% accuracy. This means they were almost perfect at guessing which students would finish and which would leave.
• The Runner-Up: Random Forest was also very strong at 98.7%.
• The Others: KNN and the more complex SVM version were good too, but slightly less accurate (around 96-97%).

What the "Crystal Ball" Told Them

When the researchers looked at why the winning models made their predictions, they found some interesting patterns:

• Visa Type: Students with a standard "Student Visa" were much more likely to finish their studies.
• Nationality: The data showed that students from Pakistan had a higher likelihood of dropping out compared to others in this specific dataset.
• Grades: The number of subjects a student failed was a huge red flag. The more they failed, the higher the risk they would leave.
• Course Choice: Students in the "Master of IT" program were more likely to stay and finish.

The Safety Check

The researchers were careful about privacy. They removed all names and personal details from the data, treating the students like anonymous numbers. They also noted a potential issue: if the data doesn't have enough examples of certain groups (like specific genders or nationalities), the computer might get biased. They suggested checking for this in the future.

The Bottom Line

This study proves that by using simple math and computer models, universities can act like proactive gardeners. Instead of waiting for a student to fail, they can use data to spot the warning signs early. The study found that simple models (Logistic Regression and Linear SVM) are actually the best tools for this job, offering a 99% chance of correctly identifying students who need help before they decide to leave.

The researchers suggest that in the future, schools could add even more details—like how much a student participates in class or their homework habits—to make these predictions even sharper, but for now, the current data is already doing a fantastic job.