wenmin-wu/tabular-popularity-fallback-recommendation

Popularity Fallback Recommendation

Overview

Personalized models often produce fewer than K recommendations for inactive or new users. Fill remaining slots with the most popular items from a recent time window. This simple fallback significantly boosts MAP@K by ensuring every user gets a full recommendation list.

Quick Start

import pandas as pd

def get_popular_items(transactions, days=7, top_k=12):
    cutoff = transactions["t_dat"].max() - pd.Timedelta(days=days)
    recent = transactions[transactions["t_dat"] >= cutoff]
    return recent["article_id"].value_counts().head(top_k).index.tolist()

def fill_recommendations(user_recs, popular, k=12):
    """Pad user's rec list with popular items up to k."""
    seen = set(user_recs)
    for item in popular:
        if len(user_recs) >= k:
            break
        if item not in seen:
            user_recs.append(item)
            seen.add(item)
    return user_recs[:k]

Workflow

1. Compute global top-K popular items from recent time window
2. For each user, generate personalized recommendations
3. If list has fewer than K items, pad with popular items (skip duplicates)
4. Ensure every user has exactly K recommendations

Key Decisions

• Recency window: 7 days typical; shorter for fast-changing catalogs
• Deduplication: Never recommend an item already in the personalized list
• Ordering: Popular items always go after personalized ones
• Cold-start: For users with zero history, the full list is popular items

References

• H&M Personalized Fashion Recommendations (Kaggle)
• Source: recommend-items-purchased-together-0-021