Build models for housing and wind power prediction
Overview
This question evaluates ML system design and applied modeling competencies, including data assumptions, preprocessing, feature engineering, model selection, validation strategies, temporal leakage handling, and end-to-end pipeline construction for a binary housing-affordability classifier and a time-series regression for wind-farm power prediction.
Two-Part Machine Learning Take-Home
Part 1 — Binary Classification: "Can Buy" vs "Cannot Buy"
Given applicant and market data, design a binary classifier to predict whether an applicant can buy a house (labels: can buy, cannot buy). Specify the following:
- Data assumptions and label definition.
- Preprocessing steps (missing values, outliers, leakage guards, encoding, scaling).
- Feature design (engineered affordability features, credit/market features).
- Model choice and rationale (including interpretability/calibration if applicable).
- Validation strategy (splits, class imbalance handling, threshold selection).
- Evaluation metrics and decision thresholding (including business-aligned metrics).
Part 2 — Regression: Wind-Farm Power Output
You are provided three files: train.csv, test.csv, sample_submission.csv. The target column is power output in train.csv. For each record in test.csv, predict power output and produce a submissions.csv with exactly two columns and a header: id, power output.
Describe your end-to-end approach:
- Data checks and exploratory analysis (schema, leakage risks, target sanity checks).
- Feature engineering (physics-aware, statistical, temporal, and interaction features).
- Time-aware validation to avoid leakage (rolling/sliding windows; lag construction).
- Model selection and tuning approach (baselines through advanced models).
- Metrics (primary/secondary) and error analysis.
- Training and inference pipeline, including how you would generate submissions.csv.
Solution
Login required