Content-Based Recommender Systems
Recommending from what an item actually is
Collaborative filtering and matrix factorization both ignore what an item actually contains — they only look at who interacted with it. Content-based recommendation flips that: it represents each item as a feature vector built from its actual attributes (genre, text description, cast, ingredients, tags, specs) and recommends items similar to ones a user already liked, based on feature similarity rather than other users' behavior.
If you loved a dense, slow-paced science fiction novel, a content-based system looks for other books sharing similar themes, pacing, vocabulary, or genre tags — regardless of whether anyone else who read it also read the recommended book.
Building item profiles
The core building block is the item profile — a feature vector describing the item. For text-heavy items (articles, product descriptions, book summaries), the classic approach is TF-IDF (term frequency–inverse document frequency): words common within an item but rare across the whole catalog get higher weight, since they're the words that best distinguish that item.
For structured catalogs (movies with genre/director/cast fields, products with category/brand/price), features are typically one-hot or multi-hot encoded and concatenated. Modern systems increasingly use learned embeddings from language or vision models instead of hand-built features — the same underlying idea, with richer representations.
Once every item has a feature vector, a user profile is built as a weighted combination (often just an average) of the profiles of items that user has liked, rated highly, or purchased.
Note
Content-based vs. collaborative filtering
| Dimension | Content-based | Collaborative filtering |
|---|---|---|
| Data required | Item features/metadata | User-item interaction history |
| New item cold start | No problem — features exist immediately | Severe — no interactions yet |
| New user cold start | Still needs some liked items to build a profile | Severe — no history to compare against |
| Discovers unexpected connections | Limited to feature similarity, less serendipitous | Can surface non-obvious cross-genre links |
| Depends on good features/metadata | Yes — garbage features, garbage recommendations | No — works directly from behavior |
Note
Hybrid recommender systems
Production systems (Netflix, Spotify, Amazon) virtually never rely on a single pure approach. Common hybridization strategies:
Weighted hybrid — compute scores from both a content-based model and a collaborative model, then combine them with a weighted sum or learned blending model.
Switching hybrid — use content-based recommendations for new items/users (cold start), and switch to collaborative filtering once enough interaction data accumulates.
Feature augmentation — feed content features directly into a collaborative model (e.g. as side information in a matrix factorization model, or as inputs to a neural recommender), letting one model learn from both signal types jointly.
Cascade — use one method to generate a candidate shortlist, then rerank with the other.
What's next
Hybrid systems exist largely to patch the weakest point of collaborative filtering: brand-new users and items with no history. The Cold Start Problem in Recommender Systems lesson looks at this challenge directly — what to recommend when you genuinely know nothing yet.
I build these systems professionally.
Whether it's a RAG pipeline, analytics migration, or AI workflow — let's talk.