Coderush 2026 — Economic Class Classification

Coderush 2026 · ML Module — Bag-level multiclass economic class classification.
Primary metric: Macro F1 · Tie-breaker: Middle Class F1

---

Problem Overview

Each sample in the dataset is a bag — a variable-sized collection of individuals described by demographic and economic attributes (education, occupation, capital gains, hours worked, etc.). The task is to predict a single economic class label for the entire bag:

Label	Class	Description
0	lower	Individuals in the lower economic tier
1	middle	Individuals in the middle economic tier
2	upper	Individuals in the upper economic tier

This is a Multiple-Instance Learning (MIL) problem — supervision is at the bag level only; no per-individual labels are provided.

---

Dataset

Split	Rows	Bags
Train	16,776	3,360
Test	1,981	400

Class distribution (train bags):

• middle: 1,260
• upper: 1,120
• lower: 980

---

Approach

1. Bag Aggregation

Raw rows are grouped by bag_id and collapsed into a single feature vector per bag using statistical aggregations (mean, std, min, max, count, sum, nunique) over all numeric and categorical columns. This converts the MIL setup into a standard tabular classification task.

• Train shape after aggregation: (3360, 176)
• Test shape after aggregation: (400, 175)
• 174 engineered features used for training

2. Multi-Seed × 5-Fold Cross-Validation

To reduce variance and produce reliable out-of-fold (OOF) predictions for blending and threshold tuning, we train across 5 random seeds × 5-fold stratified CV for each model.

Seed	OOF Macro F1	Fold scores
42	0.7089	0.7042 · 0.7383 · 0.6983 · 0.6876 · 0.7162
123	0.7205	0.7069 · 0.7200 · 0.7362 · 0.7203 · 0.7192
777	0.7239	0.7202 · 0.7103 · 0.7209 · 0.7192 · 0.7489
2024	0.7210	0.7409 · 0.7300 · 0.7029 · 0.7317 · 0.6995
31415	0.7228	0.7409 · 0.6986 · 0.7104 · 0.7421 · 0.7219

3. Model Ensemble

Three gradient boosting frameworks were trained and blended:

Model	Role
LightGBM	Primary learner (70% weight)
XGBoost	Secondary learner (30% weight)
CatBoost	Explored; 0% in final blend

Blend weights were searched over OOF predictions:

Best OOF F1: 0.7127  →  LGBM=0.7, XGB=0.3, CB=0.0

4. Decision Threshold Tuning

Per-class decision thresholds were tuned on the combined OOF probability predictions to maximise Macro F1. This gave the biggest single improvement in the pipeline:

Class	Threshold
lower	0.25
middle	0.25
upper	0.55

Pre-threshold  OOF Macro F1: 0.7127
Post-threshold OOF Macro F1: 0.7410  (+0.028)

5. Test Predictions

lower: 136  |  middle: 212  |  upper: 52
Submission: submission_v3.csv

---

Results

Metric	Value
OOF Macro F1 (pre-threshold)	0.7127
OOF Macro F1 (final)	0.7410
Competition placement	Top 5

---

Repository Structure

.
├── coderush26_v2.ipynb               # Main solution notebook (run end-to-end)
├── Coderush-26-ML-Train.csv          # Training data (16,776 rows)
├── Coderush-26-ML-test.csv           # Test data (1,981 rows, no labels)
├── submission_v3.csv                 # Final Kaggle submission
├── Coderush26_ML_ProblemStatement.pdf # Official problem statement
└── README.md

---

Quick Start

Install dependencies:

pip install lightgbm xgboost catboost scikit-learn pandas numpy

Run the notebook:

Open coderush26_v2.ipynb and run all cells. The notebook expects the CSV files in the same directory (update paths in the data-loading cell if needed). It will output:

• A saved model checkpoint (.pkl / .joblib)
• submission_v3.csv with bag-level predictions

---

Key Takeaways

• Bag aggregation is the critical preprocessing step — collapsing variable-length bags into fixed-size feature vectors is what makes standard tabular ML applicable here.
• Threshold tuning on OOF probabilities gave +2.8 F1 points with zero additional training cost.
• Multi-seed ensembling provides more stable OOF estimates for blend-weight and threshold search than a single seed.
• CatBoost added noise in this setting despite being a strong baseline on many tabular tasks — the blend search correctly assigned it 0% weight.
• Judges reward simplicity: a well-tuned gradient boosting ensemble outperformed more complex approaches while remaining fully interpretable.

---

Competition Details

Item	Detail
Competition	Coderush 2026 — ML Module
Task	Multiclass bag-level classification
Primary metric	Macro F1
Tie-breaker	Middle Class F1
Platform	Kaggle
Deadline	9 May 2026 · 08:00 AM PKT
Presentation day	10 May 2026