Reduce overfitting under constraints
Overview
This question evaluates a candidate's competency in machine learning for mitigating overfitting in tabular regression under production latency constraints, testing knowledge of regularization, model complexity control, feature engineering, validation strategies, and experimental design.
Boston Consulting Group
Oct 13, 2025, 9:49 PM
Data Scientist
Technical Screen
Machine Learning
0
0
Reduce Overfitting Under Latency Constraints (Tabular Regression)
Context (assumed)
- You have a tabular regression model with a large generalization gap: train RMSE = 4,000 vs validation RMSE = 9,500.
- You cannot collect more data. Online inference must remain under 20 ms p95 on production hardware.
- Assume the baseline is a gradient-boosted decision tree model (e.g., LightGBM/XGBoost) serving on CPU.
Task
Choose and prioritize three interventions to reduce overfitting. For each intervention:
- Explain the mechanism for why it reduces overfitting.
- Provide a concrete experiment plan with hyperparameter grids, metrics to monitor, stopping criteria, and how you will establish statistically significant improvement.
Options to consider
- L1/L2/elastic-net regularization (and expected effect on coefficients/leaf weights)
- Early stopping with patience
- Architecture or tree-depth reduction
- Feature selection / target encoding with smoothing
- Data augmentation suitable for tabular data
- K-fold cross-validation with stratification
- Bagging vs boosting
- Leakage checks
Deliverables
- A prioritized list of three selected interventions with mechanisms.
- An experiment plan covering: hyperparameter grids, monitoring metrics, stopping criteria, significance testing, and how you will enforce the 20 ms p95 latency constraint.
Solution
Login required
Show