Content-based Filtering

Content-based Recommendations · Attribute-based Filtering

Definition

In content-based filtering each item is described as an attribute vector. The user profile is the weighted average of the attributes of the items they viewed or bought, and the recommendations are the items with the highest cosine similarity to that profile.

How the algorithm works

1. Catalogue: each item becomes an attribute vector
   Nike Air Max: {brand: Nike, category: sneakers,
                  colour: white, price: 120, material: mesh}

2. User profile: the weighted average of the attributes of viewed items
   User_123: {brand: Nike ×0.6, Adidas ×0.3, category: sneakers ×0.8,
              price band: 90–160}

3. Recommendation: the items most similar to the profile
   → cosine similarity(User_123, Item_A) = 0.91 ✓
   → cosine similarity(User_123, Item_B) = 0.34 ✗

Content-based versus collaborative filtering

Dimension	Content-based	Collaborative
Works from	Item attributes	Behaviour of similar users
Cold start	✓ Works	✗ Fails
New items	✓ Works	✗ No data
Serendipity	✗ Predictable	✓ Finds the unexpected
Precision at high data volume	Medium	High

The hybrid approach

Most production systems combine algorithms:

New visitor (< 5 events):      content-based (100%)
Returning (5–20 events):       content-based (60%) + collaborative (40%)
Active (20+ events):           collaborative (70%) + content-based (30%)
New item (< 10 sales):         content-based (80%) + popularity (20%)

The hybrid removes the weakness each algorithm carries on its own.

Where it is used in e-commerce

Similar items on a product page — the closest attribute matches to the item being viewed
New arrivals — items with no interaction history that collaborative filtering cannot place
Niche categories — low-traffic sections where behavioural data never reaches critical mass
Explainable recommendations — because you looked at X, a message that only attribute logic can
justify honestly

The practical rule is to treat content-based filtering as the floor of a recommender rather than its
ceiling: it guarantees a relevant answer in every situation, and collaborative models add the
precision on top once the data is there.

Frequently asked questions

How does content-based filtering differ from collaborative filtering?

Content-based works from item attributes and one user's profile, independent of other shoppers. Collaborative filtering works from the behaviour patterns of similar users — people who bought this also bought that. Content-based is better at cold start and for niche items; collaborative is more precise once data volume is high.

When is content-based filtering the better choice?

At the cold start of a new user, when there is too little data for collaborative filtering. For new items with no sales history. For niche or custom products with unique attributes. And when recommendations need to be explainable: recommended because you were looking at Nike running shoes.

What is the weakness of content-based filtering?

The serendipity problem: the algorithm only recommends things similar to what has already been seen. Someone who searched for sneakers will keep seeing sneakers, with no unexpected but relevant finds. Diversity comes from combining it with collaborative filtering in a hybrid.

Gravity Field