Recommender Systems

By the end, you can

• Describe the user-item rating matrix and explain why sparsity is the core challenge.
• Contrast content-based and collaborative filtering, and identify the key misconception about what CF looks at.
• Trace neighborhood CF through a small rating matrix to predict a missing rating.
• Explain what latent factors are and how matrix factorization trains them via gradient updates on known ratings.
• Compute a predicted rating as the dot product of a user vector and an item vector.
• Compute cosine similarity between two unit-length taste vectors.
• Distinguish explicit from implicit feedback and explain why missing entries should not automatically be treated as dislikes.
• Identify both flavors of cold start and describe the hybrid strategies that address them.
• Calculate precision@k for a given ranked list and explain why ranking metrics are preferred over MSE.
• Describe how a two-tower model works and why pre-computed item embeddings enable real-time ranking at scale.
• Explain the recommender feedback loop, define popularity bias and filter bubbles, and name strategies to mitigate them.