Skip to content
Tutorials Center
My learning

Topics

Lesson 2 of 38 ยท Getting Started

Types of Recommender Systems

Compare popularity, content-based, collaborative, knowledge-based, context-aware and hybrid recommenders, and learn when each one fits.

There is no single "recommendation algorithm". There is a family of approaches, each using different information and failing in different ways. Knowing the map makes the rest of the course easier, because every technique later on is a refinement of one of these ideas.

This lesson names the main types, shows the same tiny catalog being recommended three different ways, and gives you a rule of thumb for picking one.

The main families

Five families of recommenders

Popularity Same top list for everyone
Content-based Items like what you liked
Collaborative People like you liked it
Knowledge-based Rules and constraints

Hybrids combine several.

1 / 2

Show the most-viewed, best-rated or fastest-rising items to everyone. It is trivial to build, hard to beat for brand-new users, and the baseline you should compare every fancier model against.

Content-based filtering

Describe every item by its attributes (genre, director, text, price, tags). Build a profile of what the user liked, then recommend items whose attributes match the profile. It needs no data about other users, so it works for new items, but it tends to stay inside what the user already knows. Chapter 3 covers it.

Collaborative filtering

Ignore item attributes completely. Use only the pattern of who interacted with what. If two users liked many of the same items, one user's other likes become suggestions for the other. It finds surprising connections, but it fails when there is no history. Chapters 4 and 5 cover neighborhood methods and matrix factorization.

Knowledge-based and constraint-based

For rarely bought, high-stakes items (a house, a laptop, a loan) there is little history. The system asks for requirements ("budget under 1,000, at least 16 GB RAM") and filters or ranks by rules. Interaction data is not required.

Context-aware and sequence-aware

Time, location, device and the previous few actions matter. "What to watch tonight on the TV with the family" differs from "what to watch on the phone during a commute". Session-based and sequence models treat a user's recent actions as an ordered series.

Hybrid

Real products blend families: collaborative signals for people with history, content features for new items, popularity as a fallback, rules for business constraints. Chapter 6 shows hybrid designs.

The same catalog, three recommenders

Here is a tiny catalog of six films with genres, plus who watched what. We recommend for a user who has watched "Space Saga" using three methods.

Python Popularity vs content vs collaborative

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

All three use a different signal. Popularity looks at counts, content at genres, collaborative at co-viewing. The lists overlap but are not identical, and that is exactly why hybrids exist.

Choosing a family

Python A rule-of-thumb chooser

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

This function is a deliberate simplification, but the questions are the right ones: do I have interaction history, do I have good item features, and do new items arrive constantly?

Other distinctions you will hear

  • Memory-based vs model-based. Memory-based methods (neighborhoods) keep the raw data and compute at query time. Model-based methods (factorization, neural nets) learn compact parameters first.
  • Explicit vs implicit data. Star ratings versus clicks and purchases; see Explicit and Implicit Feedback.
  • Pointwise, pairwise, listwise. How a model's training objective treats the ranking: score each item alone, compare pairs, or optimize the whole list.
  • Item-to-item vs user-to-item. "Related to this product" pages versus a personalized home page.

Always start with a popularity baseline. If a complex model cannot beat "show the top 10 most watched", something is wrong with the model or the data.

Check your understanding

Quick check ยท 4 questions

of 4 answered You scored / 4 ยท perfect!

  1. 1. Which family needs no information about other users?

    Content-based filtering compares item attributes with the user's own profile, so other users are not needed.
  2. 2. A new shop launches with zero purchase history but detailed product specs. What is the most sensible start?

    Without interactions, collaborative methods have nothing to learn from. Item attributes and stated requirements still work.
  3. 3. Why do production systems often use hybrids?

    Collaborative signals are strong with history, content features help new items, and popularity is a fallback. Combining them reduces failure modes.
  4. 4. What does 'memory-based' mean for neighborhood methods?

    Memory-based (also called instance-based) methods store the data itself rather than learning compact parameters.

Summary

  • Main families: popularity, content-based, collaborative, knowledge-based, context-aware, and hybrids.
  • Content-based uses item attributes; collaborative uses only the interaction pattern.
  • Cold start decides much of the choice: no history favors content, knowledge or popularity.
  • Memory-based methods compute from raw data; model-based methods learn parameters.
  • Always keep a popularity baseline.

Next: The Recommender Pipeline: Candidates and Ranking