wenmin-wu/tabular-implicit-als-collaborative-filtering

Implicit ALS Collaborative Filtering

Overview

For implicit feedback (purchases, clicks — no ratings), factorize the sparse user-item interaction matrix into low-rank latent factors using ALS. The implicit library provides a GPU-accelerated implementation. Recommendations come from dot-product similarity between user and item factor vectors.

Quick Start

import implicit
import numpy as np
from scipy.sparse import coo_matrix

def build_interaction_matrix(df, user_col, item_col):
    users = df[user_col].astype("category")
    items = df[item_col].astype("category")
    data = np.ones(len(df))
    return coo_matrix((data, (users.cat.codes, items.cat.codes)))

coo = build_interaction_matrix(transactions, "customer_id", "article_id")
model = implicit.als.AlternatingLeastSquares(
    factors=100, iterations=15, regularization=0.01
)
model.fit(coo.tocsr())

# Recommend for a user
ids, scores = model.recommend(user_id, coo.tocsr()[user_id], N=12)

Workflow

1. Build sparse COO matrix from transaction logs (user × item)
2. Convert to CSR format for efficient row slicing
3. Fit ALS model with tuned factors/iterations/regularization
4. Generate top-N recommendations per user via model.recommend
5. Batch inference in chunks (2000 users) for memory efficiency

Key Decisions

• Factors: 50-200; higher captures more patterns but risks overfitting
• Temporal filtering: Train on recent window (7-21 days) to capture current trends
• Validation: MAP@K on a held-out recent time window, not random split
• GPU: implicit supports GPU; 10x faster for large matrices

References

• H&M Personalized Fashion Recommendations (Kaggle)
• Source: h-m-implicit-als-model-0-014