Content-Based Filtering

Best for: Recommending similar items Aliases: Content-Based

How it works

Builds a user profile $h_u$ by aggregating the feature vectors $x_i$ of items the user has liked, then scores new items by similarity to that profile, e.g. $\hat{r}_{ui}=\cos(h_u,x_i)$ or a learned scorer $f_\theta(x_i)^\top h_u$. Because it relies on item features (text, tags, attributes) rather than other users’ ratings, it handles item cold-start well. The trade-off is over-specialisation: it rarely surfaces items outside the feature pattern the user already engaged with.

When to use

Recommendation from item features when interaction data is thin; for cold-start of long-tail items.

Watch out

Over-specializes to items similar to what the user already likes; quality is bounded by feature engineering.

Common fields

Retail · jobs · real estate · articles