Business Context
TalentLens, a recruiting platform screening 2M+ technical applications per year, wants an automated model to identify resumes that demonstrate credible deep learning experience. Recruiters use this score to prioritize candidates for ML interviews, but false positives are costly because they waste interviewer time.
Dataset
You are given a labeled resume-screening dataset built from historical recruiter decisions and post-interview outcomes.
| Feature Group | Count | Examples |
|---|
| Structured profile | 14 | years_experience, highest_degree, num_ml_projects, publications_count |
| Skills & keywords | 120 | cnn, pytorch, tensorflow, transformers, computer_vision |
| Resume text embeddings | 256 | sentence embedding dimensions from resume summary/projects |
| Project metadata | 10 | github_stars, kaggle_rank, deployed_models_count |
| Target | 1 | credible_deep_learning_candidate |
- Size: 48K candidate resumes, 400 features after preprocessing
- Target: Binary label indicating whether the candidate passed recruiter screening for deep learning-focused roles
- Class balance: 18% positive, 82% negative
- Missing data: 12% missing in project metadata, 6% missing in structured profile fields, no missing in embeddings
Success Criteria
A good solution should achieve strong ranking quality for recruiter triage: PR-AUC >= 0.55, recall >= 0.80 at precision >= 0.45, and calibrated probabilities suitable for thresholding by hiring teams.
Constraints
- Inference must score a resume in <50 ms in an online screening API.
- Recruiters need some interpretability at the feature-group level.
- The model must be retrained monthly as skill trends change.
- Budget favors a compact tabular/deep hybrid model over large transformer fine-tuning.
Deliverables
- Build a binary classification pipeline to predict credible deep learning experience from resume data.
- Explain model choice versus simpler baselines such as logistic regression or gradient boosting.
- Handle class imbalance, missing values, and mixed feature types.
- Evaluate the model with threshold-free and threshold-based metrics.
- Propose how you would deploy, monitor, and retrain the model in production.
