Screen Internship Candidates with ML

Business Context

TalentMatch, a recruiting platform processing ~120K internship applications per hiring cycle, wants a model to predict whether a candidate will pass the first-round technical screen. Recruiters need a practical baseline model to prioritize reviews while keeping the process explainable and fair.

Dataset

You are given historical candidate records from internships and student projects. Each row represents one application.

• Size: 48K applications, 33 features
• Target: Binary — passed first-round screen (1) vs not passed (0)
• Class balance: Moderately imbalanced, 28% positive and 72% negative
• Missing data: 12% missing in GPA, 18% missing in coding_test_score, sparse skill indicators for some resumes

Success Criteria

A good solution should achieve ROC-AUC >= 0.82, F1 >= 0.68, and recall >= 0.75 for the positive class at an operational threshold. The model should also provide interpretable feature importance for recruiters.

Constraints

• Batch scoring only; daily inference on up to 10K new applications
• Recruiters need understandable drivers behind predictions
• Training must fit on a standard CPU machine within 30 minutes
• Avoid leakage from post-screen outcomes or manually added recruiter notes

Deliverables

1. Build a classification pipeline to predict first-round screen pass/fail.
2. Explain model choice versus a simpler baseline such as logistic regression.
3. Handle missing values, categorical variables, and class imbalance appropriately.
4. Evaluate with cross-validation and a held-out test set.
5. Report feature importance and threshold tradeoffs for recruiter operations.