CASE 1 Default
The factory settings from the software library. No tuning at all. This is the baseline.
Analogy: wearing the outfit exactly as it came out of the packaging.
Reproduction of Demirtürk & Harunoğlu (2025)
Can a computer predict a student’s GPA — and how well can we reproduce a published answer?
We rebuilt a research paper’s experiments from scratch: 10 prediction models, each tuned 3 ways, then compared our results with the original, number by number.
02 The problem
Predict a GPA from twelve facts about a student
In plain wordsGiven things like study time and absences, guess the student’s GPA, a number from 0 to 4.
What is regression?
Regression means predicting a number (here, GPA) rather than choosing a category (like pass/fail).
The model studies thousands of past students, learns which inputs move the number up or down, then applies that pattern to a student it has never seen.
The 12 inputs
Target: GPA (0–4).
Data leakage, and why we avoided it
Data leakage is when the model is accidentally handed the answer. The dataset also has GradeClass (A–F), but that letter is calculated from GPA. Keeping it would be like letting a student copy the answer key.
So GradeClass was removed, along with StudentID (a label with no meaning). The paper made the same choice.
Result: a fair test where the model only sees information it would really have in advance.
03 Dataset & exploration
What the data says before any model runs
In plain wordsOne thing dominates: students who miss more classes get much lower GPAs.
Kaggle Students Performance dataset: no missing values and no duplicate rows.
Absences vs GPA: −0.92
A correlation near −1 means “as one goes up, the other goes down almost in a straight line.” Absences is by far the strongest signal.
Everything else is a whisper
Weekly study time, parental support and tutoring have weak positive effects. Age, gender, ethnicity, sports and music have almost none.
04 Pipeline
From raw spreadsheet to a fair comparison
In plain wordsClean the data, hide a test set, train and tune on the rest, then grade every model on the hidden students.
…or click / tab to any step.
Show the original flow diagram from the report
05 The three cases
Every model gets tuned three ways
In plain wordsModels have dials, called hyperparameters. We compare leaving the dials alone, trying every setting, and evolving good settings.
CASE 2 Grid Search
Write down a short list of values for each dial and try every combination, scoring each with cross-validation (repeated practice exams on the training data only).
- 24–180 combinations per model
- Thorough, but cost explodes as dials are added
Analogy: trying on every outfit in the wardrobe.
CASE 3 Genetic Algorithm
Start with random settings. The best ones “survive”, mix with each other (crossover), occasionally get a random tweak (mutation), and the next generation is usually better.
- Mutation rate 0.05 · elitism 10%
- Selection · crossover · mutation
Analogy: breeding better candidates over generations.
06 The 10 models
Ten ways to make the same prediction
In plain wordsDifferent “thinking styles” for the same task. Click a card to see the best settings we and the paper found.
Hyperparameter text: teal = ours, blue-grey = paper. Full tables with match flags are in section 8d.
07 Results explorer
Which model predicts GPA best?
In plain wordsR² is the share of GPA variation a model explains (1.0 is perfect). RMSE and MAE are the typical size of the miss, in GPA points, so lower is better.
Tuning case
Show
Metric
The chart library could not load (are you offline?). The sortable table below has all the same numbers.
Sortable table
★ = best in that column (highest R² / CV, lowest error). The paper reports no cross-validation scores, so its CV column is “—”. Paper values: Tables 3, 6, 8. Ours: report Tables 1–3 and notebook.
08 Paper vs my implementation
Same experiment, independent rebuild
In plain wordsWe repeated the paper’s experiment and checked whether we got the same answers. Mostly yes; where we differ, we say why. This is a reproduction study, not a critique of the authors.
Setup side by side
In plain wordsWe matched the paper on data, split, models and metrics. The biggest deliberate difference is the smaller genetic-algorithm budget.
| Aspect | Base paper | My implementation |
|---|
Do our R² values match the paper’s?
In plain wordsFor 27 of 30 experiments, our score is within 0.016 of the paper’s. The 3 big gaps are all cases where the paper’s own result collapsed.
≈ equal |Δ| ≤ 0.005small 0.005–0.016large > 0.016Δ = Ours − Paper (test-set R²)
Each cell shows Δ. Colour and sign/label carry the meaning.
| Model | Case | Paper R² | Ours R² | Δ (Ours − Paper) | Agreement |
|---|
Which model won each case?
In plain wordsSVR wins everywhere in our run; the paper agrees for two cases but crowns XGBoost for the genetic algorithm.
SVR error (RMSE) — paper vs ours
| Case | Paper | Ours |
|---|
Which settings did each search pick?
In plain wordsWhere two searches land on the same dial positions, it suggests both found the same sweet spot.
| Model | Paper | Ours | Match |
|---|
Why are some results so different?
Likely explanations from our analysis, not proven facts. We have not re-run the paper’s exact code, so each point below is a hypothesis to verify.
Likely cause Bagging Regressor paper ≈ 0.14 / 0.11 → ours ≈ 0.93
The paper’s tables list max_features = 1. In scikit-learn, an integer 1 means “each model may look at exactly one feature”, while the decimal 1.0 means “look at 100% of the features”.
max_features=1 → 1 of 12 columns per tree
max_features=1.0 → all 12 columns per tree
max_features=1.0 → all 12 columns per tree
Predicting GPA from a single input would give very poor R². Our run used 1.0 and did not collapse. To be verified by re-running with the integer.
Plausible HGBR with GA paper 0.748 → ours 0.936
The paper’s GA picked learning_rate = 0.01 with absolute_error loss, a very slow learner that probably under-trained with the default number of iterations. Our GA picked learning_rate = 0.231.
Observed GA is not automatically better than Grid Search
With a much smaller GA budget (20 × 5 vs 100 × 10), our GA scored below Grid Search for GBM, HGBR and XGBoost. “GA is better” is not guaranteed; it depends on the search budget.
Hypothesis Small default-case gaps GBM 0.9384 vs 0.9463 · XGBoost 0.9305 vs 0.9344
Probably due to library-version differences in default settings.
What each study concludes
In plain wordsWe agree on the winners (SVR, GBM, XGBoost). We disagree that Bagging fails, and we add that a plain straight-line model does just as well.
Base paper
- SVR, GBM and XGBoost combine high explanatory power with low error.
- Ensemble methods, especially GBM and XGBoost, are the most stable.
- Bagging (and partly HGBR) should not be preferred on this dataset.
- Recommends SVR and GBM as balanced choices.
- Future work: Bayesian Optimization, Random Search, other datasets, deep learning, extra variables like motivation.
Our reproduction
- Confirms SVR (best in all three cases); GBM, XGBoost and HGBR near the top (0.94–0.95).
- Does not reproduce the Bagging failure (R² ≈ 0.93).
- Tuning helps weak models most (DT +0.025–0.029, KNN +0.014, AdaBoost +0.012, XGBoost +0.013 with Grid Search); strong models barely change.
- GA does not always beat Grid Search.
- Adds: plain Linear/Ridge/Lasso regression reaches R² = 0.9532, matching the best tuned model, because GPA is almost linear in Absences.
Our contributions beyond the paper
In plain wordsExtra analyses that explain why the models behave as they do.
Things we noticed while reproducing
Our observations from reading the paper; please double-check against the original. These made exact reproduction harder, so we followed the paper’s intent and documented our choices.
09 Key findings
Seven things to remember
In plain wordsThe short version, if you read nothing else.
10 Limitations & future work
What this study can’t tell you
In plain wordsGreat scores on one tidy dataset don’t guarantee the same on real schools.
Limitations
- Single dataset, likely synthetic (almost perfectly linear in Absences).
- No feature scaling, to match the paper.
- Reduced GA budget: 20 × 5 instead of 100 × 10.
- Near-linear target limits how much models can differ.
Future work
- Run the GA at the paper’s full 100 × 10 budget for a fair GA-vs-Grid comparison.
- Try Bayesian Optimization and Random Search.
- Test on other datasets from different institutions.