Collaborative filtering recommends items based on user-item interaction patterns.
Core idea: users who agreed in the past may agree in the future.
User-Based Collaborative Filtering
User-based collaborative filtering finds users similar to the target user, then recommends items those similar users liked.
Steps:
- Find users similar to the target user.
- Identify items those users liked.
- Recommend items the target user has not seen.
Example: “Users like you also bought…”
Item-Based Collaborative Filtering
Item-based collaborative filtering finds items similar to the user’s liked items.
Steps:
- Calculate similarity between items.
- For items the user liked, find similar items.
- Recommend similar unseen items.
Example: “Customers who bought this also bought…”
Similarity Metrics
Collaborative filtering often uses Distance Metrics And Similarity Measures such as cosine similarity, Pearson correlation, or Jaccard similarity.
Advantages
- can discover unexpected items
- does not require hand-crafted item features
- improves with more interaction data
Disadvantages
- suffers from cold start
- sparse user-item matrices are hard to learn from
- popular items can dominate
- explanations can be less direct than content-based methods
Matrix Factorization is a common model-based form of collaborative filtering.