Recommender Systems I: Content-Based and Collaborative Filtering

Introduction

Recommender systems help users discover items: movies, products, songs, posts, jobs, restaurants, or documents.

Two classic families are:

• Content-based recommendation.
• Collaborative filtering.

Modern systems often combine both, but the distinction is still useful.

Content-Based Recommendation

Content-based systems recommend items similar to what a user already liked.

They rely on item features:

• Movie genre.
• Product category.
• Text description.
• Tags.
• Price.
• Brand.
• Embeddings.

The system builds a user profile from previously liked or consumed items, then recommends items with similar features.

Pros

• Works for new items if item features exist.
• Easier to explain.
• Does not require many users.

Cons

• Can become repetitive.
• Depends on feature quality.
• Struggles to discover surprising items outside the user’s history.

Collaborative Filtering

Collaborative filtering uses behavior patterns across users and items.

The core idea:

Users who behaved similarly in the past may like similar items in the future.

The data is often represented as a user-item matrix:

rows: users
columns: items
values: ratings, clicks, purchases, watch time, or implicit feedback

User-Based Collaborative Filtering

Find users similar to the target user, then recommend items those similar users liked.

Item-Based Collaborative Filtering

Find items similar to items the user liked, then recommend those items.

Item-based methods are often more stable because item similarities can change more slowly than user preferences.

Similarity

Common similarity measures:

• Cosine similarity.
• Pearson correlation.
• Jaccard similarity.
• Dot product of embeddings.

The right measure depends on whether feedback is explicit ratings, implicit actions, binary events, or dense embeddings.

Cold Start

Cold start happens when the system lacks history.

New users:

• Ask onboarding questions.
• Use popularity or trending items.
• Use contextual signals.
• Use content-based recommendations.

New items:

• Use item metadata.
• Use text or image embeddings.
• Explore with limited traffic.

Hybrid systems often exist because pure collaborative filtering struggles with cold start.

Evaluation

Offline metrics:

• Precision@k.
• Recall@k.
• nDCG.
• Mean reciprocal rank.
• Coverage.
• Diversity.

Online metrics:

• Click-through rate.
• Conversion.
• Watch time.
• Retention.
• Revenue.
• User satisfaction.

A recommender can improve clicks while hurting long-term trust. Always use guardrail metrics.

Closing

Content-based methods use item meaning. Collaborative filtering uses collective behavior. Strong recommender systems usually combine both and evaluate beyond a single ranking metric.