Recommendation Systems

Recommendation systems predict what users might like based on their past behavior, preferences, and patterns from similar users. Unlike supervised learning with explicit labels, recommendations learn from implicit signals—what users click, view, purchase, or rate.

What Makes Recommendations Different

Personalization: Every user gets different recommendations tailored to their preferences. The same item might be perfect for one user and irrelevant for another.

Implicit feedback: Most systems don't have explicit ratings. Instead, they infer preferences from behavior—clicks, purchases, watch time, skips.

Cold start: New users or new items have no history. The system must make recommendations with minimal or no data.

Diversity vs. accuracy tradeoff: Highly accurate recommendations might all be similar (filter bubble). Good systems balance accuracy with discovery.

Scale: Millions of users x millions of items = trillions of possible recommendations. Systems must be efficient.

Types of Recommendation Systems

Collaborative Filtering

Learn from patterns across many users. "Users who are similar to you also liked..."

User-based: Find similar users, recommend what they liked.

• Pros: Discovers new items, captures trends
• Cons: Scalability issues, cold start for new users

Item-based: Find similar items to what the user liked.

• Pros: More stable, scalable, explainable
• Cons: Less discovery, cold start for new items

Matrix Factorization: Decompose user-item matrix into latent factors.

• Pros: Handles sparsity well, scalable, good accuracy
• Cons: Cold start problem, harder to explain

Content-Based Filtering

Recommend items similar to what the user liked based on item features.

How it works: If you liked science fiction movies, recommend other science fiction movies. Uses item metadata, descriptions, tags, or features.

Pros:

• No cold start for new items (if you have their features)
• Recommendations are explainable
• Works with few users

Cons:

• Needs good item features/descriptions
• Filter bubble—can't discover outside interests
• Cold start for users with no history
• Over-specialization

Hybrid Systems

Combine collaborative and content-based approaches to get the best of both.

Strategies:

• Weighted: Average predictions from both methods
• Switching: Choose one method based on context
• Feature combination: Use CF and content features together in one model
• Cascade: Use one method to filter, then another to rank

Pros: Overcomes limitations of individual approaches, better cold start handling

Cons: More complex, harder to tune, requires both interaction data and content features

Knowledge-Based

Use explicit rules and constraints. Common for complex, infrequent purchases like real estate, cars, insurance.

Example: "Show me 3-bedroom houses under $500k within 10 miles of downtown"

Sequential/Session-Based

Predict the next item in a sequence. Captures short-term intent and temporal patterns.

Examples: Next video to watch, next item to add to cart, next song in playlist

Models: Recurrent neural networks, transformers (BERT4Rec), Markov chains

Association Rules

"Frequently bought together" patterns. Based on transaction co-occurrence.

Example: Customers who buy diapers also buy baby wipes

Key Concepts

Explicit vs Implicit Feedback

Explicit: Direct ratings, likes/dislikes, thumbs up/down

• Clear signal of preference
• Sparse—most users don't rate most items
• Can be biased (only engaged users rate)

Implicit: Clicks, purchases, watch time, page views

• Abundant and continuous
• Ambiguous—did they like it or were curious?
• Reflects actual behavior, not stated preference

Cold Start Problem

New user: No interaction history

• Solutions: Ask preferences onboarding, demographic-based, popular items, content-based

New item: No user interactions yet

• Solutions: Content-based recommendations, show to active users first, "New arrivals" section

New system: No users or items

• Solutions: Bootstrap with popular items, external data, knowledge-based rules

Sparsity

Most users interact with very few items. The user-item matrix is 99%+ empty.

Challenges: Hard to find similar users/items with enough overlap

Solutions: Matrix factorization (learns latent patterns), hybrid methods, dimensionality reduction

Evaluation Metrics

Precision@K

Of the K recommendations, what fraction did the user actually interact with?

Example: Recommend 10 movies, user watches 3 -> Precision@10 = 0.3

Use: Measures recommendation accuracy. Higher is better.

Recall@K

Of all items the user liked, what fraction appear in top K recommendations?

Example: User likes 20 movies total, 5 appear in top 10 -> Recall@10 = 5/20 = 0.25

Use: Measures coverage of user interests. Higher is better.

NDCG (Normalized Discounted Cumulative Gain)

Ranking quality metric that accounts for position. Items at the top matter more.

Interpretation: 0-1 scale, higher is better. Considers both relevance and ranking.

Use: When order matters. A relevant item at position 1 is better than at position 10.

Hit Rate@K

What fraction of users have at least one relevant item in top K?

Example: 80% of users find something they like in top 10 -> Hit Rate@10 = 0.8

Use: Measures if the system works for most users.

Coverage

What fraction of items ever get recommended?

Interpretation: Low coverage = most recommendations go to popular items (less diversity)

Use: Detect filter bubbles and popularity bias.

Diversity

How different are the recommended items from each other?

Measurement: Average pairwise dissimilarity between recommended items

Use: Avoid showing 10 nearly identical items. Balance with accuracy.

Serendipity

Surprising recommendations that users enjoy but wouldn't find themselves.

Not just accurate—also unexpected and delightful. Hard to measure.

Common Challenges

Popularity Bias

Popular items dominate recommendations. Niche items rarely surface.

Why it happens: Popular items have more data, safer predictions, reinforcement loops

Solutions: Normalize by popularity, boost under-recommended items, diversity metrics

Filter Bubble

Users only see recommendations similar to past behavior. No discovery.

Solutions: Inject diversity, exploration, trending items, cross-domain recommendations

Position Bias

Users click top results more, regardless of relevance. This biases training data.

Solutions: Debiasing techniques, randomization, position-aware models

Feedback Loop

System recommends popular items -> they get more clicks -> become even more popular

Solutions: Exploration, randomization, decay popularity over time

Choosing an Approach

New to recommendations, simple ratings data: Start with Matrix Factorization (SVD)

Lots of users, need fast updates: Item-Based Collaborative Filtering

Have rich item descriptions/metadata: Content-Based or Hybrid

Need explainability: Item-Based CF or Content-Based (can show similar items)

Large scale, many features: Deep learning (BERT4Rec for sequences)

E-commerce, "bought together": Association Rules

Semantic understanding needed: Embeddings (sentence transformers)

Sequential behavior matters: BERT4Rec or session-based models

Cold start is critical: Hybrid or Content-Based

Practical Considerations

Data Requirements

Minimum: User-item interactions (who interacted with what)

Better: Ratings, timestamps, purchase amounts, dwell time

Best: + item features (descriptions, categories, tags), user features (demographics, preferences)

Scalability

Millions of users/items: Matrix Factorization, Item-Based KNN with approximate neighbors

Real-time recommendations: Precompute item similarities, use fast lookups, cache user profiles

Daily/weekly batch: Any method works, focus on accuracy

Business Metrics

Optimize for clicks: Precision@K, CTR

Optimize for sales: Revenue per recommendation, conversion rate

Optimize for engagement: Watch time, session length, return rate

Optimize for discovery: Coverage, diversity, serendipity

Don't just maximize accuracy—align with business goals.

A/B Testing

Always validate recommendations in production:

• Control (baseline) vs. treatment (new model)
• Measure actual user behavior, not offline metrics
• Watch for unintended consequences (filter bubbles, bias)

Example Workflow

1. Understand Your Data

What interactions do you have? (views, purchases, ratings)
How sparse is the data? (% of user-item pairs with interactions)
Do you have timestamps? (enables sequential models)
Do you have item features? (enables content-based)

2. Start Simple

Baseline: Popularity ranking
Next: Item-Based CF or Matrix Factorization
Evaluate with offline metrics (Precision@K, NDCG)

3. Handle Cold Start

New users: Popular items + onboarding questions
New items: Content-based recommendations initially
Monitor cold start metrics separately

4. Improve Incrementally

Add content features -> Hybrid model
Add timestamps -> Sequential model
Tune hyperparameters
Ensemble multiple models

5. Deploy and Monitor

A/B test in production
Monitor business metrics
Track diversity and coverage
Detect and mitigate bias
Update regularly

Relationship to Other Tasks

Clustering: Group users or items to understand segments. Use clusters as features in recommendations.

Classification: Predict explicit user preferences (will they like this? yes/no). Less common than regression.

Regression: Predict ratings (1-5 stars). Used in explicit feedback systems.

Association Analysis: Find items frequently bought together. Complementary to collaborative filtering.

Embeddings: Learn user and item representations. Foundation for many modern recommender systems.

Recommendations often combine multiple techniques. A sophisticated system might use clustering to find user segments, embeddings to represent items, and transformers to model sequences—all working together.

Next Steps

Ready to build recommendations? The training guide covers:

• 9 different algorithms from simple to advanced
• When to use each approach
• How to configure parameters
• Handling cold start and sparsity
• Evaluation and tuning strategies

Start with Matrix Factorization or Item-Based KNN, then experiment with hybrid and sequential models as your data and requirements grow.