How is content-based recommendation related to information retrieval?

Very closely: the user profile plays the role of a query, items play the role of documents, and recommending the best items is essentially ranking documents by similarity to the query. Many content-based methods reuse retrieval representations such as tf-idf and cosine similarity.

What is over-specialization in content-based recommendation?

Because the method only recommends items similar to what a user already liked, it can keep suggesting near-duplicates and miss novel or surprising items the user might enjoy. This lack of serendipity is a key reason content-based methods are often combined with collaborative filtering.

Content-Based Recommendation

Content-based recommendation suggests items whose features resemble those of items a user has previously liked, building a profile of the user's interests from item content.

Definition

Content-based recommendation predicts a user's interest in an item by comparing the item's content features with a profile of the user's preferences inferred from items they have previously interacted with, recommending items most similar to the profile.

Scope

This topic covers recommendation that relies on item descriptions and a learned user profile: representing items by their content features, building user profiles from rated or consumed items, and matching profiles to candidate items using similarity or supervised learning. It addresses the close kinship with information retrieval, where the user profile acts like a standing query, and the characteristic strengths and limitations such as over-specialization. It excludes collaborative methods that use other users' behavior.

Core questions

How are items represented by content features for recommendation?
How is a user profile constructed from the items a user has liked?
How are candidate items scored against the user profile?
How does content-based recommendation relate to information retrieval and filtering?
Why does content-based recommendation tend toward over-specialization?

Key concepts

item content features
user profile
tf-idf and keyword profiles
profile-item similarity
supervised preference learning
over-specialization
explainability of recommendations
new-item handling

Key theories

Profile-as-query matching: By representing items with content features and the user with a profile aggregated from liked items, content-based recommendation reduces to retrieval-style similarity matching, where the profile functions like a persistent query against the item catalog.
Strengths and over-specialization: Content-based methods can recommend new and niche items and explain suggestions by features, but because they only surface items similar to the user's history, they risk a narrow, over-specialized set lacking serendipity.

Clinical relevance

Content-based recommendation drives suggestions for articles, products, music, and videos where item descriptions are rich, and it handles new items well because it does not depend on others' ratings. It is widely combined with collaborative methods to offset each approach's weaknesses.

History

Content-based recommendation grew directly out of information filtering and retrieval in the 1990s, treating a user profile as a long-standing query. Pazzani and Billsus's overview and the later state-of-the-art survey by Lops and colleagues consolidated representation and profile-learning techniques, and the approach remains a standard component, especially within hybrid systems.

Key figures

Michael Pazzani
Daniel Billsus
Pasquale Lops
Giovanni Semeraro

Seminal works

pazzani2007
lops2011

Frequently asked questions

How is content-based recommendation related to information retrieval?: Very closely: the user profile plays the role of a query, items play the role of documents, and recommending the best items is essentially ranking documents by similarity to the query. Many content-based methods reuse retrieval representations such as tf-idf and cosine similarity.
What is over-specialization in content-based recommendation?: Because the method only recommends items similar to what a user already liked, it can keep suggesting near-duplicates and miss novel or surprising items the user might enjoy. This lack of serendipity is a key reason content-based methods are often combined with collaborative filtering.