Item-Based Collaborative Filtering
What you'll learn: How to compute similarities between items and use a user's own ratings to predict what they'll think of new items.
The Core Idea
Instead of finding similar users (like User-Based Collaborative Filtering), Item-Based Collaborative Filtering finds similar items. The logic: if you liked "The Matrix" and it's similar to "Inception" (because many users rated both highly), you'll probably like "Inception" too.
The key shift is computing item-to-item similarities once, then using your own ratings to predict your interest in new items. This approach often scales better because item relationships tend to be more stable than user preferences.
How It Works
- Compute item similarities: For every pair of items, calculate how similarly users have rated them using metrics like cosine similarity or Pearson correlation (covered earlier)
- Store the similarity matrix: This becomes your item-to-item lookup table
- Predict ratings: When predicting user u's rating for item i:
- Find items similar to i that user u has already rated
- Weight those ratings by the similarity scores
- Compute a weighted average as the prediction
Why This Matters
Items change less frequently than user preferences. A movie's "similarity neighborhood" remains relatively stable, while a user's tastes might shift. This means you can precompute item similarities offline and serve predictions quickly at runtime.
It also handles the cold-start problem differently: new users can get recommendations immediately based on just a few ratings, because item relationships are already established.
Key Takeaway: Item-Based Collaborative Filtering flips the perspective—instead of "users like you enjoyed this," it says "items like the ones you rated highly include this," offering better scalability and stability for many real-world recommender systems.