Recommendation Systems

Core Problem

• Predict: Will user U like item I?
• Rank: Which items should we show first?
• Filter: Remove irrelevant or inappropriate content
• Discover: Help users find new interests

Business Impact

• 35% of Amazon revenue from recommendations
• 80% of Netflix viewing from recommendations
• Increased user engagement and retention
• Reduced choice overload

User-Based CF

$$\text{sim}(u,v) = \frac{\sum_{i \in I_{uv}} (r_{ui} - \bar{r}_u)(r_{vi} - \bar{r}_v)}{\sqrt{\sum_{i \in I_{uv}} (r_{ui} - \bar{r}_u)^2} \sqrt{\sum_{i \in I_{uv}} (r_{vi} - \bar{r}_v)^2}}$$

• Find users similar to target user
• Recommend items liked by similar users
• Works well for niche items

Item-Based CF

• Find items similar to user's liked items
• More stable than user-based
• Easier to explain recommendations
• Better for large user bases

Challenges

• Cold Start: New users/items have no history
• Sparsity: Most user-item pairs unobserved
• Scalability: Computation grows with users/items
• Popularity Bias: Favors popular items

Similarity Metrics

• Cosine Similarity: Angle between vectors
• Pearson Correlation: Linear relationship
• Jaccard Similarity: For binary data

Feature Extraction

• Text: TF-IDF, word embeddings, topics
• Images: CNN features, color histograms
• Audio: Spectrograms, MFCC features
• Metadata: Genre, director, price, brand

User Profile Building

• Aggregate features from liked items
• Weight by ratings or implicit feedback
• Update profile over time
• Handle preference drift

Advantages

• No cold start for new items
• Transparent recommendations
• Domain knowledge integration
• No sparsity issues

Limitations

• Limited content analysis
• Over-specialization risk
• Feature engineering required
• Cold start for new users

SVD and NMF

$$R \approx U \Sigma V^T$$ where $U$ are user factors, $V$ are item factors

• SVD: Handles missing values with modifications
• NMF: Non-negative factors, interpretable
• ALS: Alternating Least Squares optimization

Neural Collaborative Filtering

• Replace dot product with neural networks
• Learn complex user-item interactions
• Combine with deep features

Autoencoders

• Encode user preferences into latent space
• Reconstruct missing ratings
• Handle sparse data naturally
• Variational AE for uncertainty

Embedding Techniques

• Item2Vec: Skip-gram for items
• User2Vec: Learn user representations
• Graph Embeddings: User-item graph structure

Hybrid Strategies

• Weighted: Combine algorithm scores
• Switching: Choose algorithm by context
• Cascade: Sequential refinement
• Feature Combination: Unified model

Context-Aware Systems

• Time of day, season, location
• Device type and platform
• Social context and mood
• Current activity or task

Sequential Models

• RNNs/LSTMs: Model temporal patterns
• Session-based: Short-term preferences
• Next-item prediction: Immediate needs

Exploration vs Exploitation

• Multi-armed bandits: Balance known vs new
• Thompson sampling: Bayesian approach
• ε-greedy: Random exploration

Offline Metrics

• RMSE/MAE: Rating prediction accuracy
• Precision@K: Relevant items in top-K
• Recall@K: Coverage of relevant items
• NDCG: Normalized discounted cumulative gain
• AUC: Area under ROC curve

Beyond Accuracy

• Diversity: Variety in recommendations
• Novelty: Surprising but relevant items
• Coverage: Catalog item distribution
• Serendipity: Unexpected discoveries

Online Evaluation

• A/B Testing: Compare algorithms live
• Click-through Rate: User engagement
• Conversion Rate: Purchase/consumption
• Session Length: User retention

Business Metrics

• Revenue per user
• Customer lifetime value
• User satisfaction scores
• Return visit frequency

Scalability Challenges

• Real-time inference: Sub-100ms latency
• Batch processing: Precompute recommendations
• Distributed computing: Spark, Hadoop clusters
• Caching strategies: Redis, CDN integration

Data Pipeline

• Streaming user interactions
• Feature engineering pipelines
• Model retraining schedules
• A/B testing infrastructure

Bias and Fairness

• Popularity bias: Over-recommend popular items
• Filter bubbles: Narrow user exposure
• Demographic bias: Unfair treatment by group
• Feedback loops: Rich get richer

Privacy Considerations

• Federated learning approaches
• Differential privacy techniques
• User data anonymization
• GDPR compliance requirements

Implementation Best Practices

• Start Simple: Basic CF before complex models
• Measure Everything: Comprehensive A/B testing
• Handle Cold Start: Have fallback strategies
• Explain Recommendations: Build user trust
• Monitor Bias: Ensure fair recommendations

Common Pitfalls

• Optimizing only for accuracy
• Ignoring temporal dynamics
• Over-engineering early systems
• Not handling data sparsity

Emerging Trends

• Graph Neural Networks: Complex relationship modeling
• Reinforcement Learning: Long-term user satisfaction
• Causal Inference: Understanding recommendation effects
• Federated Learning: Privacy-preserving recommendations
• Multimodal Systems: Text, image, audio integration

Success Metrics

• User engagement and retention
• Revenue and conversion impact
• Diversity and catalog coverage
• User satisfaction surveys