Content-Based Recommendation

Definition

Content-Based Recommendation

Content-based recommendation (a.k.a. content-based filtering) recommends items to a user based on the content/attributes of the items the user has previously liked, rather than on the collective interaction data of other users. Each item is described by a content profile (features extracted from text, audio, images, metadata, etc.), and each user is described by a user profile built from the content of items they interacted with. Recommendation reduces to finding items whose content profile best matches the user’s profile.

It is one axis of the content-vs-collaborative-vs-hybrid paradigm: where Collaborative Filtering uses who liked what, content-based filtering uses what the item is.

Intuition

"More of what you already liked"

If a user repeatedly listens to acoustic folk tracks, describe those tracks by their content (genre tags, audio features, lyrics), average them into a profile of the user’s taste, and then recommend other tracks whose content is closest to that profile. No other users are needed — the system reasons purely about item attributes and the individual’s own history. This is why content-based methods shine in domains like news (recency-driven, content-rich text) where fresh items have no interaction history yet.

Mathematical Formulation

Represent every item $i$ by a content feature vector ${\bar{c}}_i\in {\mathbb{R}}^d$ (e.g., a TF-IDF vector over the words/tags describing the item). Build the user profile ${\bar{p}}_u$ as a (rating-weighted) aggregate of the content of items the user has interacted with:

${\bar{p}}_u=\frac{{\sum }_{i\in I_u}{r}_{ui}{\bar{c}}_i}{{\sum }_{i\in I_u}|r_{ui}|}$

Score a candidate item by the cosine similarity between user profile and item content:

$\hat{s}(u,i)=\cos ({\bar{p}}_u,{\bar{c}}_i)=\frac{{\bar{p}}_u\cdot {\bar{c}}_i}{\Vert {\bar{p}}_u\Vert \Vert {\bar{c}}_i\Vert }$

where:

• ${\bar{c}}_i$ — content feature vector of item $i$ (e.g., TF-IDF weights over the terms describing item $i$; $d$ = vocabulary/feature size)
• $I_u$ — set of items user $u$ has interacted with (the user’s history)
• $r_{ui}$ — feedback signal of $u$ on $i$ (explicit rating, or implicit $1$); positive feedback pulls the profile toward ${\bar{c}}_i$
• ${\bar{p}}_u$ — user profile, a weighted centroid of liked items’ content in the same feature space
• $\hat{s}(u,i)$ — match score; items are ranked by $\hat{s}$ and the top-K are returned

This profile-then-cosine recipe is the classic linear instance (the Rocchio Algorithm applied to recommendation). More generally, the matcher can be any classifier/regressor $f_{\theta }({\bar{c}}_i)$ trained per user on their history (e.g., naive Bayes, k-NN, or a neural model), predicting how relevant item content is to that user.

Key Properties / Variants

• No reliance on other users. Scores depend only on the target user’s own history + item content, which gives it complementary strengths to Collaborative Filtering.
• Item cold start is solved. A brand-new item with zero interactions can still be recommended the moment its content features are known — crucial for news/music where the catalog churns fast.
• User cold start persists. A new user with an empty $I_u$ has no profile, so the method cannot personalize yet.
• Feature engineering is the bottleneck. Quality depends entirely on how well ${\bar{c}}_i$ captures what matters. Text → TF-IDF / Word Embeddings; audio/image/video → learned embeddings. Neural encoders let the model ingest heterogeneous content (text, images, audio) without hand-crafted features.
• Over-specialization / filter bubble. Because it only retrieves “more of the same,” it has low Diversity/Serendipity — it struggles to surface genuinely novel items outside the user’s established profile.
• Transparent & explainable. Recommendations are justifiable via shared content features (“recommended because it shares genre/keywords with X”).
• Hybrid use. Usually combined with CF in a Hybrid Recommendation to get the best of both: content for cold items, collaborative signal for popular ones.

Algorithm: Content-Based Recommendation
────────────────────────────────────────
# 1. Item profiles (offline)
for each item i in catalog:
    c_i ← extract_features(i)        # e.g. TF-IDF over text, or learned embedding
 
# 2. User profile (per user u)
p_u ← weighted_mean({ r_ui * c_i  for i in history(u) })
 
# 3. Scoring & ranking (online, for target user u)
for each candidate item i not in history(u):
    score[i] ← cosine(p_u, c_i)       # or f_θ(c_i) for a trained matcher
return top-K items by score[i]

Connections

• Paradigm sibling of: Collaborative Filtering (interaction-based) — content-based is the content axis
• Combined with CF in: Hybrid Recommendation
• Linear instance is essentially: Rocchio Algorithm (profile + cosine matching)
• Feature representations: TF-IDF, Word Embeddings, Vector Space Model
• Addresses item side of: Cold Start Problem
• Weakness motivates: Diversity, Serendipity, Novelty (beyond-accuracy)
• Ranked output evaluated with: Recall, MRR, NDCG

Appears In

• RS-L01 - Course Overview & Introduction