wenmin-wu/cv-multilabel-cooccurrence-correlation-eda

Overview

Multi-label problems hide a label-graph: classes that always appear together, classes that mutually exclude, and classes that anchor others. A vanilla df.corr() over the indicator columns is dominated by the empty-row baseline and shows nothing. Restricting to rows with n_labels > 1, then computing correlation, surfaces real co-occurrence. Pair this with per-rare-class conditional counts (P(label_j | label_i) for each rare i) to find which common classes systematically accompany each rare one. The output of this analysis seeds feature-engineering decisions: which classes to merge into a shared head, which to weight in a hierarchical loss, which to enforce as mutually-exclusive at post-processing.

Quick Start

import seaborn as sns
import matplotlib.pyplot as plt

# 1) Restrict to multi-label rows and compute correlation
multi = train_labels[train_labels['number_of_targets'] > 1]
corr = multi[class_cols].corr()
sns.heatmap(corr, cmap='RdYlBu', vmin=-1, vmax=1, square=True)
plt.title('Multi-label co-occurrence correlation')
plt.show()

# 2) Conditional counts: which classes accompany a rare anchor?
def cooccur_with(anchor, df, class_cols):
    sub = df[df[anchor] == 1][class_cols].sum(axis=0)
    return sub[sub > 0].sort_values(ascending=False)

print(cooccur_with('Rods & rings', train_labels, class_cols))
print(cooccur_with('Microtubule ends', train_labels, class_cols))

Workflow

1. Add an n_labels column counting positive labels per row
2. Filter to rows with n_labels >= 2 for the correlation pass
3. Compute and plot the correlation heatmap; cluster rows/cols with clustermap to surface groups visually
4. For each rare class, list the top-K classes it co-occurs with — note anchors
5. Use the discovered groups to: (a) add weighted multi-task heads, (b) introduce a hierarchical co-occurrence loss term, (c) write post-hoc rules ("if X, then never predict Y")
6. Re-run after model training to compare prediction co-occurrence vs. ground-truth — large divergence flags a missing constraint

Key Decisions

• Filter to multi-label rows before correlating: empty-row mass makes everything look anti-correlated otherwise.
• Use clustermap not heatmap for >15 classes: hierarchical clustering surfaces the groups automatically.
• Conditional counts beat correlation for rare classes: with 5 positives, correlation is unstable; raw counts are interpretable.
• Don't infer causation: co-occurrence reflects biology / dataset construction, not predictive direction.
• Compare to predicted co-occurrence post-training: divergence tells you if the model is honoring the label structure.
• Generalizes: any multi-label problem (audio tagging, ICD-10 medical coding, multi-genre recsys) benefits from the same EDA pass.

References

• Protein Atlas - Exploration and Baseline