wenmin-wu/tabular-multiclass-to-binary-collapse

Multiclass-to-Binary Collapse

Overview

When auxiliary labels provide finer-grained categories than the competition's binary target (e.g., disease subtypes A/B/C/D vs healthy/sick), training on the multiclass target gives the model more signal to learn decision boundaries. At prediction time, sum the probabilities of all non-baseline classes to get the binary positive probability. This typically improves binary classification by 0.5-2% because the model learns distinct patterns per subtype instead of lumping them together.

Quick Start

import numpy as np
from sklearn.model_selection import StratifiedKFold

# y_multi has subtypes: ['A', 'B', 'C', 'D', 'E']
# 'A' = healthy, 'B','C','D','E' = different conditions
# Binary target: class_0 = 'A', class_1 = everything else

# Train on multiclass target
model.fit(X_train, y_multi_train)

# Predict multiclass probabilities
probs = model.predict_proba(X_test)  # shape: (n, 5)

# Collapse: class_0 = P(A), class_1 = P(B) + P(C) + P(D) + P(E)
class_0_prob = probs[:, 0]  # baseline class
class_1_prob = probs[:, 1:].sum(axis=1)  # all non-baseline

submission['class_0'] = class_0_prob
submission['class_1'] = class_1_prob

Workflow

1. Map auxiliary labels to a multiclass target (subtypes)
2. Train classifier on the multiclass target
3. Predict multiclass probabilities
4. Sum all non-baseline class probabilities for the positive class
5. Submit binary probabilities

Key Decisions

• Baseline class: Must be index 0 in model.classes_ — verify alignment
• Oversampling: Use the multiclass target for stratified oversampling (preserves subtype balance)
• CV stratification: Stratify folds by the multiclass target, not the binary
• When to use: Only when subtypes have genuinely different feature patterns

References

• ICR Identify Age