Dual-Agency Housing Challenge 🏠

2nd Place Solution — Regression with Distribution Shift Handling

Overview

This repository contains my solution for the APPLAI ’26 Dual-Agency Housing Challenge, where the goal was to predict median house values from two heterogeneous data sources with severe distribution shifts.

The focus of this work was not only model accuracy, but also robust data preprocessing, distribution-shift detection, and reproducible ML engineering.
The final solution achieved 2nd place on the private leaderboard.

---

Key Contributions

• Detected and analyzed distribution shifts between merged data sources.
• Designed custom preprocessing transformers to handle:
  • Unit and scale mismatches
  • Bimodal feature distributions
  • Outliers and extreme values
• Built a K-Fold neural network training pipeline with ensembling.
• Saved preprocessing artifacts and trained models for full reproducibility.
• Performed systematic experimentation to improve generalization.

---

Methodology

1. Data Exploration & Shift Detection

• Exploratory data analysis to compare feature distributions across agencies.
• Identification of inconsistent units and bimodal behaviors in multiple features.

2. Preprocessing Pipeline

Custom preprocessing components were implemented to stabilize training:

• UnitAligner: Aligns features with inconsistent measurement units.
• BimodalSplitter: Separates and normalizes bimodal feature distributions.
• Clipper: Limits extreme outliers to reduce training instability.
• Robust scaling and missing-value handling.

All preprocessing steps are fitted on training data only and saved as artifacts.

3. Model Training

• Neural networks built with TensorFlow / Keras.
• K-Fold cross-validation to reduce variance.
• Ensembling across folds for final predictions.
• Evaluation using Mean Squared Error (MSE).

4. Experiments

• Baseline vs enhanced preprocessing
• Log-transformed targets
• Loss function variations
• Seed-based ensembling
• Stacking and averaging strategies

---

Results

• Private Leaderboard Rank: 🥈 2nd Place
• Private Score (MSE): 0.23875

Submission file:

• submission_enhancements.csv

---

Technologies Used

• Python
• TensorFlow / Keras
• Scikit-learn
• Joblib
• NumPy / Pandas
• Git

---

Reproducibility

• Fixed random seeds for experiments.
• Saved preprocessing pipelines and trained models using joblib.
• Clear separation between training, validation, and inference steps.

---

Notes & Limitations

• The solution is optimized for robustness under distribution shifts rather than minimal model complexity.
• Neural networks were chosen for flexibility; tree-based models were explored but underperformed under shifts.
• Further improvements could include uncertainty estimation and automated shift detection.

---

Author

Menna Thabet
Computer Science & AI Student