Early Student Dropout Prediction

Team of 3 · BINUS · Streamlit app + course research paper

Can dropout risk be flagged at enrollment, before first-semester grades exist?

Read the profile
Estimate the risk
Compare to the threshold

The app's default profile

Marital status: Single
Study program: Biofuel Production Technologies
Previous qualification: Secondary education
Mother's education: Secondary education
Father's education: Secondary education
Displaced student: No
Special needs: No
Gender: Female
Age at enrollment: 18
International: No

34.0%Estimated dropout probability

Below the threshold: not flagged.

Walkthrough of real model output. Random Forest, live app, 4 Oct 2026.

My role: Trained the app's model; charts and editing for the paper.
Team: 3 students: Saladin Zhalifunnas Ahfar, Michael Hendrilie, Valentino Allen Prasetyo. Supervisors: Rifqi Alfinnur Charisma, Henry Lucky
Stack: scikit-learn, XGBoost, Streamlit

Problem

Universities usually see dropout risk only after first-semester grades arrive. We tested how early that risk can be flagged.

Data: the UCI “Predict Students’ Dropout and Academic Success” dataset (Realinho et al., 2022). We removed the “Enrolled” class, leaving 3,630 records: 2,209 Graduate and 1,421 Dropout.

Approach

The project has two parts, and their numbers differ because the setups differ. The app uses 10 enrollment-time features and a threshold tuned for recall. The paper uses only pre-acceptance variables, an 80:20 split and SMOTE on the training data.

Part A: EarlyDrop app

A Streamlit screening tool. We compared five pipelines (Random Forest, Extra Trees, SVM with an RBF kernel, logistic regression and gradient boosting) and chose Random Forest, with the decision threshold set to 0.40 for recall.

It uses no semester data, so nothing from after enrollment leaks into the prediction. The code has unit tests.

Part B: Research paper

“Early Student Dropout Prediction Using Pre-Acceptance Variables with Gaussian Naive Bayes and XGBoost.” Course paper (unpublished), Research Methodology in Computer Science (COMP6696), BINUS, June 2026.

We compared Gaussian Naive Bayes and XGBoost in two feature settings: Background-Only and Background-Application.

EarlyDrop app · example profile, live app captured 4 Oct 2026

Results

EarlyDrop app

High recall has a cost. With precision at 0.537, roughly 46% of flagged students in the test set did not drop out.

Research paper

Gaussian Naive Bayes reached recall of 0.94 to 0.97 but precision below 0.40, so it over-flags. Pre-acceptance data gives a useful but moderate early signal.

EarlyDrop app, test set (Random Forest, threshold 0.40)
Metric	Value
Recall	0.838
Precision	0.537
F1	0.655
ROC-AUC	0.775
Accuracy	0.654

Research paper, XGBoost with Background-Application features
Metric	Value
Accuracy	0.733
Precision	0.651
Recall	0.683
F1	0.667

Bar chart of the ten most important XGBoost features in the Background-Application setting — Fig. 2. Top-10 XGBoost feature importance (Background-Application). Chart by Allen.

Distribution of predicted dropout probabilities for Graduate and Dropout students — Fig. 3. XGBoost prediction probability distribution. Chart by Allen.

What I did

Paper (CRediT roles): Investigation, Visualization, and Writing: review and editing. I produced the research visualizations: the model comparison, the Top-10 XGBoost feature-importance chart (Fig. 2) and the prediction-probability distribution (Fig. 3).

App: model training and threshold tuning (commits are under Saladin's account because training ran on his machine).

Limitations

EarlyDrop is decision support. Every flag needs human review before anyone acts on it.

Both parts use one public dataset, so the results may not hold at other universities.

Pre-acceptance data alone gives a moderate signal. The paper's best F1 is 0.667.

Problem

Approach

Part A: EarlyDrop app

Part B: Research paper

Results

EarlyDrop app

Research paper

What I did

Limitations

Links

EarlyDrop app

Research paper

Next project