Content-Based Recommendation

Content-based recommendation scores items from their attributes: text, tags, categories, prices, image embeddings, or metadata. It is useful when collaborative filtering is sparse, especially for cold-start items that have content but no interactions.

Profiles from item features

If $x_i$ is an item feature vector and $I_u^{+}$ are items the user liked, a simple user profile is

$$p_u=\frac{1}{|I_u^{+}|}\sum _{i\in I_u^{+}}{x}_i,$$

with score

$$s(u,j)=\frac{p_u^{\top }{x}_j}{\Vert p_u\Vert \Vert x_j\Vert }.$$

The feature vector might be TF-IDF, a learned embedding, or structured metadata.

Worked example

If a user liked items with feature vectors $(1,1,0,0)$ and $(1,0,1,0)$, the profile is their average:

$$p_u=(1,0.5,0.5,0).$$

Item	Feature vector	Already liked?	Cosine to profile	Interpretation
0	$(1,1,0,0)$	yes	filtered	Used to build the profile.
1	$(1,0,1,0)$	yes	filtered	Used to build the profile.
2	$(0,0,1,1)$	no	0.289	Shares one profile feature.
3	$(0,1,0,1)$	no	0.289	Shares a different profile feature.

Items 2 and 3 tie because each overlaps with half of the learned profile. Hybrid recommenders combine this with behavioral signals when both are available.

Caveats

Content recommenders can overspecialize because they look for more of what the profile already contains. Feature quality matters: weak tags or biased embeddings produce weak recommendations. Content similarity also does not prove user intent; final systems still need ranking, diversity, and online validation.

References

• scikit-learn documentation: cosine_similarity
• Adomavicius and Tuzhilin, 2005, Toward the Next Generation of Recommender Systems

Nav

Section — Recommendation Systems and Personalization

← Explicit Versus Implicit Feedback Collaborative Filtering →