wenmin-wu/tabular-weighted-position-decay-ensemble

Weighted Position Decay Ensemble

Overview

When combining ranked lists from multiple recommendation models, score each item as model_weight / (position + 1). Items appearing in multiple lists accumulate scores. Re-rank by total score to produce the final list. This naturally favors items ranked high by strong models while still crediting items from weaker ones.

Quick Start

def blend_recommendations(rec_lists, weights):
    """Blend multiple ranked recommendation lists.

    Args:
        rec_lists: list of lists, each a ranked recommendation
        weights: list of floats, per-model importance weight

    Returns:
        list of items sorted by blended score
    """
    scores = {}
    for recs, w in zip(rec_lists, weights):
        for rank, item in enumerate(recs):
            scores[item] = scores.get(item, 0) + w / (rank + 1)
    ranked = sorted(scores, key=scores.get, reverse=True)
    return ranked[:12]

Workflow

1. Generate ranked recommendation lists from each model
2. Assign importance weight to each model (tune on validation)
3. Score each item: accumulate weight / (rank + 1) across models
4. Sort by total score, take top-K

Key Decisions

• Decay function: 1/(rank+1) is simple; 1/log2(rank+2) is gentler (NDCG-like)
• Weight tuning: Optimize on MAP@K validation; can use Optuna
• Multi-stage: Blend top models first, then re-blend result with additional models
• Deduplication: Built-in — items accumulate scores naturally

References

• H&M Personalized Fashion Recommendations (Kaggle)
• Source: lb-0-0240-h-m-ensemble-magic-multi-blend