Streaming services recommend shows by turning viewer behavior into data and using algorithms to predict what each person is likely to watch next. Every play, pause, search, rating, skip, and rewatch can become a signal about preference. These systems matter because they help users find content in huge libraries, while also shaping what becomes popular.
A good recommendation system balances accuracy, variety, freshness, and fairness.
Key Facts
- Collaborative filtering recommends items using patterns from many users with similar behavior.
- Content-based filtering recommends items similar to what a user already liked, using features such as genre, actors, topics, or keywords.
- A user-item matrix stores interactions, where rows are users, columns are items, and entries may be ratings, clicks, or watch time.
- Matrix factorization approximates preferences as R ≈ U V^T, where U represents user factors and V represents item factors.
- Cosine similarity compares vectors with cos(theta) = (A · B) / (||A|| ||B||).
- Cold start happens when a new user or new show has too little interaction data for reliable recommendations.
Vocabulary
- Recommendation system
- A recommendation system is software that predicts which items a user may like based on data about users, items, and past behavior.
- Collaborative filtering
- Collaborative filtering is a method that recommends items by comparing a user's behavior with behavior from other users.
- Content-based filtering
- Content-based filtering is a method that recommends items similar to ones the user already watched or liked.
- Embedding
- An embedding is a list of numbers that represents a user, show, word, or image in a way a computer model can compare.
- Cold start problem
- The cold start problem is the difficulty of making good recommendations when there is little or no data about a new user or item.
Common Mistakes to Avoid
- Thinking recommendations are based only on ratings. Many services use watch time, skips, searches, device type, time of day, and other signals because most users rarely give explicit ratings.
- Confusing collaborative filtering with content-based filtering. Collaborative filtering uses patterns across users, while content-based filtering uses item features such as genre, cast, and description.
- Assuming the same show always has the same thumbnail for everyone. Services may test or personalize thumbnails because different images can attract different viewers to the same content.
- Treating a high prediction score as a guarantee that a user will watch. Recommendation scores are probabilities or rankings, and real choices are affected by mood, context, novelty, and competing options.
Practice Questions
- 1 A user watched 80 minutes of a 100 minute movie. If watch completion score = minutes watched / total minutes, what is the score?
- 2 Two show embeddings are A = (3, 4) and B = (6, 8). Compute their cosine similarity using cos(theta) = (A · B) / (||A|| ||B||).
- 3 A new viewer creates an account and has not watched anything yet. Explain why this is a cold start problem and name two pieces of information a streaming service could use to make early recommendations.