← Back to projects
Data science

Student dropout prediction

Risk model that estimates each student's dropout probability and ranks the tutoring list: who to intervene with first when resources are limited. Pilot (POC) on the UCI dataset, with the decision threshold as a business dial.

4,424
Students in dataset
65.8%
Recall with 1st semester
0.88
ROC-AUC (early version)

The Problem

Tutoring resources are limited: you cannot intervene with every student at once. The question is not just who is at risk, but who to intervene with first and how early.

  • • Minority and imbalanced class (32% dropout): accuracy alone is misleading
  • • Bounded intervention window: the early model has less data but arrives in time
  • • Every false positive is a wasted tutoring session on a student who was not going to drop out
  • • Pilot on the public UCI dataset; real institution data is requested separately

My Process

1. Understand (EDA)

Business hypotheses confronted against data before modeling.

  • ✓ Unpaid tuition: 86.6% dropout vs. 24.7% for those up to date
  • ✓ 0 courses passed in 1st semester: 79.4% dropout
  • ✓ Macroeconomic context discarded (correlations ≈ 0)

2. Prepare

Data quality audit and feature engineering.

  • ✓ Dataset with no nulls or duplicates (comes preprocessed)
  • ✓ 1st semester pass ratio as a feature
  • ✓ Cleaning recipe documented to apply to the institution's real data

3. Model

An honest ladder of models, from the naive baseline to the risk model.

  • ✓ Metric chosen before training: dropout-class recall (business decision)
  • ✓ Logistic regression (1st semester and full) vs. Random Forest
  • ✓ Operational threshold 0.53 optimized by F1

4. Decision and prioritized list

The operational deliverable: the student list ranked by risk for the tutoring team.

  • ✓ Risk bands: 87.8% real dropout in 'High' vs. 10% in 'Low'
  • ✓ The early version is chosen despite lower recall: it arrives in time to intervene
  • ✓ Each tutoring investment is justified by visible per-row signals (tuition, scholarship, passed courses)

The Solution

Why logistic regression and not Random Forest

Random Forest detects slightly more dropouts but with lower precision.

• RF: recall 0.778, precision 0.786 vs. logistic 0.722 / 0.869
• Nearly identical F1 (0.782 vs. 0.788)
• Logistic regression is interpretable: readable coefficients for the tutoring team

The threshold as a business dial

The default threshold (0.5) is not always the best for the decision.

• Raising it favors precision: fewer wasted tutoring sessions, more undetected dropouts
• Lowering it favors recall: more coverage, more false positives
• Optimizing F1: threshold 0.53 → precision 0.88, recall 0.72

Early vs. full version

The full version wins on every metric, but you have to wait a year.

• Full: recall 0.722 and AUC 0.908
• Early: 65.8% of dropouts detected in time to intervene
• Choosing which to use is a business decision, not a technical one

Results & Impact

72.2%
Recall (full version)
65.8%
Recall (early version)
0.908
ROC-AUC (full version)
0.53
Operational threshold (F1)

Key Learnings

  • ✓ Metric first: accuracy alone is misleading: the naive baseline already has 68% without detecting a single dropout.
  • ✓ Converging signal: logistic regression and Random Forest agree on the variables that matter (tuition, scholarship, 1st semester passes).
  • ✓ Business trade-off: choosing the early version with lower recall, because it arrives in time to act, is not a technical decision.

Tech Stack

Language and data

• Python 3
• Pandas
• NumPy
• UCI dataset

Modeling

• scikit-learn
• Logistic regression
• Random Forest
• 5-fold cross-validation

Visualization and app

• Plotly
• Multi-page Streamlit
• Jupyter notebooks

Interested in this kind of work?

I have more data analysis cases with models and honest validation. Let's connect.