Recommendation Engine Architecture#

Recommendations drive 35% of Amazon's revenue and 80% of Netflix watches. Behind every "you might also like" is a system balancing relevance, diversity, freshness, and latency.

Why Recommendations Matter#

Without recommendations:
  User searches → browses → maybe finds something → high bounce rate

With recommendations:
  User arrives → personalized feed → discovers items → longer sessions
  Result: 10-30% increase in engagement and conversion

The difference between a mediocre and great recommendation engine is architecture, not just algorithms.

Collaborative Filtering#

The most intuitive approach: people who liked similar things will like similar things.

User-Based Collaborative Filtering#

Find users similar to the target user, then recommend what those similar users liked.

# Simplified user-based CF
from sklearn.metrics.pairwise import cosine_similarity
import numpy as np

# User-item interaction matrix (ratings)
# Rows = users, Columns = items
ratings = np.array([
    [5, 3, 0, 1],  # User A
    [4, 0, 0, 1],  # User B
    [1, 1, 0, 5],  # User C
    [0, 0, 5, 4],  # User D
])

# Compute user similarity
user_sim = cosine_similarity(ratings)

# For User B, find most similar user → User A
# Recommend items User A rated highly that User B hasn't seen
# → Recommend Item 2 (rating 3 from User A)

Pros: No item metadata needed, captures complex patterns. Cons: Cold start for new users, doesn't scale well with millions of users.

Item-Based Collaborative Filtering#

Find items similar to what the user already liked. Amazon popularized this approach.

# Item-based CF — compute item similarity
item_sim = cosine_similarity(ratings.T)

# User liked Item 1 → find items most similar to Item 1
# Similarity scores tell us which items co-occur in ratings
# More stable than user-based (items change less than user behavior)

Why item-based often wins: Item relationships are more stable than user relationships. A user's taste shifts; the similarity between two movies doesn't.

Content-Based Filtering#

Recommend items with features similar to what the user previously liked.

# Content-based: use item features
item_features = {
    "movie_1": {"genre": "sci-fi", "director": "Nolan", "year": 2014},
    "movie_2": {"genre": "sci-fi", "director": "Villeneuve", "year": 2021},
    "movie_3": {"genre": "comedy", "director": "Gerwig", "year": 2023},
}

# User watched movie_1 and liked it
# TF-IDF or embedding on features → movie_2 is most similar
# Recommend movie_2

Pros: No cold start for items (features available immediately), transparent reasoning. Cons: Over-specialization (filter bubble), can't discover surprising recommendations.

Hybrid Approaches#

Production systems combine multiple strategies:

┌─────────────────────────────────────────────┐
│            Hybrid Recommender               │
├─────────────────────────────────────────────┤
│                                             │
│  Collaborative ──┐                          │
│  Filtering       ├──→ Weighted    ──→ Final │
│                  │    Combination     List   │
│  Content-Based ──┤                          │
│  Filtering       │                          │
│                  │                          │
│  Popularity    ──┘                          │
│  Baseline                                   │
│                                             │
└─────────────────────────────────────────────┘

Common hybrid strategies:

• Weighted: Score = 0.6 * CF_score + 0.3 * content_score + 0.1 * popularity
• Switching: Use content-based for new users, CF once enough data exists
• Cascade: CF generates candidates, content-based re-ranks them
• Feature augmentation: CF embeddings become features for a content model

Matrix Factorization#

The breakthrough behind Netflix Prize. Decompose the sparse user-item matrix into dense latent factors.

# Matrix Factorization with Surprise library
from surprise import SVD, Dataset, Reader
import pandas as pd

# Load ratings
reader = Reader(rating_scale=(1, 5))
data = Dataset.load_from_df(
    df[['user_id', 'item_id', 'rating']], reader
)

# SVD — learns latent factors for users and items
algo = SVD(n_factors=100, n_epochs=20, lr_all=0.005)
trainset = data.build_full_trainset()
algo.fit(trainset)

# Predict: user 42's rating for item 314
prediction = algo.predict(uid=42, iid=314)
# prediction.est → 4.2 (predicted rating)

Latent factors capture hidden dimensions — a movie might score high on "cerebral sci-fi" and low on "family friendly" without anyone labeling those dimensions.

Deep Learning Recommendations#

Neural networks handle complex patterns, sequences, and multimodal data.

# Two-tower model with TensorFlow Recommenders
import tensorflow as tf
import tensorflow_recommenders as tfrs

# User tower — learns user embeddings
user_model = tf.keras.Sequential([
    tf.keras.layers.StringLookup(vocabulary=user_ids),
    tf.keras.layers.Embedding(len(user_ids) + 1, 64),
])

# Item tower — learns item embeddings
item_model = tf.keras.Sequential([
    tf.keras.layers.StringLookup(vocabulary=item_ids),
    tf.keras.layers.Embedding(len(item_ids) + 1, 64),
])

# Retrieval task — find items closest to user in embedding space
task = tfrs.tasks.Retrieval(
    metrics=tfrs.metrics.FactorizedTopK(
        candidates=items.batch(128).map(item_model)
    )
)

When to use deep learning: Large datasets (millions of interactions), sequential patterns (session-based), multimodal features (text + image + behavior).

The Cold Start Problem#

The hardest challenge: recommending for new users or new items with no interaction history.

# Multi-armed bandit for cold start (Thompson Sampling)
import numpy as np

class ThompsonSampling:
    def __init__(self, n_items):
        self.alpha = np.ones(n_items)  # successes
        self.beta = np.ones(n_items)   # failures

    def select_item(self):
        samples = np.random.beta(self.alpha, self.beta)
        return np.argmax(samples)

    def update(self, item_idx, reward):
        if reward:
            self.alpha[item_idx] += 1
        else:
            self.beta[item_idx] += 1

A/B Testing Recommendations#

You can't improve what you can't measure. Every recommendation change needs rigorous testing.

Control group (50%):  Current algorithm
Treatment group (50%): New algorithm

Metrics to track:
├── Engagement: CTR, time-on-site, pages-per-session
├── Conversion: purchases, signups, completions
├── Diversity: unique items shown, category spread
├── Novelty: how "surprising" are recommendations
└── Coverage: % of catalog that gets recommended

Pitfall: Optimizing only for CTR creates clickbait. Track downstream metrics (purchases, retention) alongside clicks.

Real-Time vs Batch Processing#

Most production systems use hybrid: batch-compute candidate sets, real-time re-rank based on session context.

Batch pipeline (nightly):
  User history → Train model → Generate top-1000 candidates per user → Store in Redis

Real-time layer (per request):
  Session context → Re-rank candidates → Apply business rules → Return top-20

Tools and Frameworks#

Production Architecture#

┌──────────────┐     ┌──────────────┐     ┌──────────────┐
│  Event Stream │────→│  Feature     │────→│  Model       │
│  (Kafka)      │     │  Store       │     │  Training    │
└──────────────┘     │  (Feast)     │     │  (nightly)   │
                      └──────────────┘     └──────┬───────┘
                                                   │
┌──────────────┐     ┌──────────────┐     ┌───────▼───────┐
│  API Gateway  │←───│  Re-Ranker   │←───│  Candidate    │
│  (response)   │     │  (real-time) │     │  Store (Redis)│
└──────────────┘     └──────────────┘     └───────────────┘

Key Takeaways#

1. Start simple — popularity and item-based CF beat complex models with small data
2. Hybrid always wins — combine multiple signals for robustness
3. Solve cold start explicitly — bandits and content features fill the gap
4. Measure everything — A/B test with downstream metrics, not just CTR
5. Batch + real-time — precompute candidates, re-rank in real time

---

Build smarter systems with codelit.io — your visual architecture companion.

Article 191 of the Codelit engineering blog series.