SAR Tutorial: Smart Adaptive Recommendations

SAR Tutorial: Smart Adaptive Recommendations#

Introduction#

SAR (Smart Adaptive Recommendations) is an item-based collaborative filtering model. It uses item co-occurrence and similarity (Jaccard, lift, etc.) to produce fast, interpretable recommendations — similar in spirit to classic “users who liked this also liked…” systems.

How SAR Works#

Architecture#

SAR builds item–item similarity from user interaction history, then scores candidate items for each user from items they have already interacted with.

Mathematical Foundation#

Recommendations combine item affinity and similarity signals over the user’s history (see Microsoft Recommenders SAR for the original formulation).

Tutorial with cr_learn#

Step 1: Import and Load Data#

from corerec.engines.collaborative import SAR
from cr_learn import ml_1m
from sklearn.model_selection import train_test_split

data = ml_1m.load()
ratings_df = data['ratings']
ratings_df = ratings_df.rename(columns={"user_id": "userID", "movie_id": "itemID"})
train_df, test_df = train_test_split(ratings_df, test_size=0.2, random_state=42)

print(f"Loaded {len(ratings_df)} ratings")

Step 2: Initialize Model#

model = SAR(similarity_type='jaccard')

print(f"Initialized SAR with {model.similarity_type} similarity")

Step 3: Train#

model.fit(train_df)

print("Training complete!")

Step 4: Predict#

sample_user = train_df['userID'].iloc[0]
sample_item = train_df['itemID'].iloc[0]
score = model.predict(sample_user, sample_item)
print(f"Predicted score: {score:.3f}")

Step 5: Recommend#

recommendations = model.recommend(user_id=sample_user, top_k=10)

print(f"Top-10 recommendations for User {sample_user}:")
for rank, item_id in enumerate(recommendations, 1):
    print(f"  {rank}. Item {item_id}")

Step 6: Save & Load#

model.save('artifacts/sar')
loaded = SAR.load('artifacts/sar')
recs = loaded.recommend(sample_user, top_k=5)
print(f"Loaded model recommendations: {recs}")

Key Takeaways#

When to Use SAR#

✅ Best for:

  • Implicit or explicit feedback with item co-occurrence signal

  • Fast baseline CF without neural training

  • Interpretable item-similarity recommendations

❌ Not ideal for:

  • Strict sequential next-item prediction (use SASRec / BERT4Rec)

  • Rich side features only (pair with content models)

Best Practices#

  1. Use userID, itemID, rating column names (or pass custom columns to fit)

  2. Try similarity_type='jaccard' or 'lift' for implicit data

  3. Save with a base path (artifacts/sar) for the safe bundle default

Further Reading#

Contents