Comparing 15 Machine Learning Classifiers for University Admissions: Why Random Forest Wins

What the researcher actually did

The author sourced a public Kaggle dataset of 400 university applicants, each with seven input features and a "chance of admission" score between 0 and 1. Anyone scoring 0.80 or above was labeled "admitted"; everyone else, "not admitted." To build a more robust sample, the dataset was bootstrapped up to 800 records, then split 80/20 into training and testing sets. Because the two outcome classes weren't evenly balanced, the SMOTE resampling method was applied before training — a standard step to stop a model from simply learning to predict the majority class every time. Fifteen classifiers were then trained and compared through Python's PyCaret library, spanning simple baselines, tree-based ensembles, boosting methods, and other standard approaches.

Why Random Forest won, and by how much

Random Forest topped the leaderboard with 95.5% accuracy, and — per the paper — essentially unchanged performance between training and test data, the standard evidence a model has learned generalizable patterns rather than memorized its training examples. Extra Trees, a close relative that adds extra randomness when splitting decision-tree branches, finished a fraction of a point behind. The two Random Forest-family models clearly separated themselves from the rest of the field: boosting variants landed in the 92-94% range, a single Decision Tree came in around 93%, and simpler linear/statistical models clustered in the 87-88% range. The linear-kernel SVM was the clear outlier at 60.8% — because a linear model can only draw a straight-line boundary between classes, and the real relationship between an applicant's features and their outcome appears to be more complex than that.

What actually predicts admission

Using Random Forest's built-in feature-importance scores, the study ranked which of the seven inputs mattered most: CGPA first at roughly 30% of total importance, GRE score second at about 22%, then TOEFL score, university rating, statement of purpose, and letters of recommendation in modestly declining order — with research experience, often assumed to be a differentiator, ranking last.

Why it matters in practice

The paper's practical pitch: a model this accurate could serve as a first-pass screening tool for admissions offices, flagging likely-admit and likely-reject applications for faster processing while leaving final judgment calls to human reviewers. The trained model was saved for exactly that kind of downstream use — as a web app, an API, or an integration into an existing admissions system.

Reader Q&A

Q: What is Random Forest, in plain terms, and why did it win?

A: It's an ensemble method that trains many individual decision trees on random subsets of the data and features, then combines their votes into one final prediction. That averaging effect makes it more accurate and less prone to overfitting than any single decision tree — which is why it outperformed both simpler models and most other ensemble methods here.

Q: How confident can we be that the model isn't just overfitting to this dataset?

A: The paper reports very similar accuracy on training data (95.5%) and previously unseen test data (also 95.5%), plus a low out-of-bag error of 3.14%. Consistent performance across training and test sets is the standard sign a model has learned generalizable patterns rather than memorized specific examples.

Q: Does a high GRE or TOEFL score guarantee admission under this model?

A: No single feature guarantees anything — the model weighs all seven factors together. But the feature-importance analysis shows GRE and TOEFL scores carry substantially more predictive weight than softer factors like the statement of purpose, with CGPA mattering most of all.

Q: Why did the linear-kernel SVM perform so much worse than the other models?

A: A linear SVM can only separate classes with a straight-line boundary. If the real relationship between an applicant's features and their outcome is more complex or non-linear — which tree-based methods can capture but a linear model cannot — a linear SVM will systematically underperform, which is what the 60.8% accuracy result suggests here.

Q: Is this dataset representative of real-world, current admissions criteria?

A: It's a single, relatively small public dataset — 400 original records, expanded synthetically to 800 — reflecting one set of historical outcomes, not a live or universal standard. The relative importance of CGPA over research experience may not hold at every institution or country, a caveat worth keeping in mind before generalizing the findings broadly.