A recommendation technique that suggests items by matching the attributes of items a user has previously engaged with — such as text features, genres, tags, or learned embeddings — against the attributes of candidate items, building a per-user preference profile in feature space; it requires no data about other users, handles new items gracefully, and offers explainable suggestions, but tends to over-specialise, recommending only items similar to what the user already knows.
Semantic Classification
Content
Definition
Content-based filtering recommends items to a user by comparing item attributes with a profile of that user’s demonstrated preferences. Each item is represented as a feature vector — classically TF-IDF weights over text, genre and metadata tags, or, in modern systems, dense embeddings from neural encoders. The user profile is an aggregate (often a weighted mean) of the vectors of items they have rated, watched, or purchased. Candidate items are scored by similarity to this profile, typically with cosine similarity or a learned relevance model, and the top-scoring items are surfaced.
The approach contrasts directly with Collaborative Filtering, which ignores item content and instead exploits patterns across the user–item interaction matrix (“users like you also liked…”). Content-based methods need no other users’ data, which gives them three practical advantages: new items can be recommended immediately from their attributes (no item cold-start), recommendations remain available for niche users with unusual tastes, and each suggestion can be explained in terms of concrete features. Their characteristic failure is over-specialisation — the system keeps recommending close neighbours of past consumption and rarely produces serendipitous discoveries, a dynamic that feeds the Filter Bubble phenomenon.
Production recommenders almost always deploy hybrids: content features handle cold-start and long-tail items, collaborative signals supply serendipity and social proof, and a ranking model blends both. Within this graph, content-based filtering is one functional component of Content Discovery pipelines and of hyper-personalisation systems.
Technical Details
A minimal pipeline has three stages. Item representation: extract features per item — TF-IDF or BM25 for text; one-hot or learned embeddings for categorical metadata; CNN, audio, or transformer embeddings for images, music, and long-form text. Profile construction: aggregate the vectors of positively engaged items, usually with recency weighting and negative feedback subtraction; Rocchio’s relevance-feedback formula is the classical formulation. Scoring: rank candidates by cosine similarity or feed profile and item vectors into a learned model (logistic regression through to two-tower neural retrieval, where one tower encodes the user history and the other the item content). Modern large-scale systems blur the line: two-tower and sequence models such as SASRec consume content embeddings as inputs, making “content-based” one signal within a learned ranking stack rather than a standalone algorithm. Evaluation uses ranking metrics (precision@k, nDCG, MAP) plus beyond-accuracy measures — diversity, novelty, and catalogue coverage — precisely because content-based systems score well on accuracy while narrowing exposure.
Current Landscape
-
Semantic IDs: content-derived discrete item tokens (RQ-VAE quantised embeddings) have become the dominant successor to random hashed item IDs; Spotify Research (September 2025) showed that embeddings fine-tuned jointly for search and recommendation give the best cross-task Semantic IDs.
-
Generative retrieval in production: YouTube and Google DeepMind’s PLUM framework (arXiv:2510.07784, October 2025) fuses multi-modal content embeddings into SID-v2 tokens and is in production for YouTube recommendations, delivering a reported +4.96% panel CTR lift for Shorts in live A/B tests — content signals now sit inside LLM-native generative recommenders rather than standalone content-based scorers.
-
LLM-as-recommender scaling: a September 2025 study (arXiv:2509.25522) found SID-based generative recommendation saturates with scale, whereas directly using LLMs as recommenders (44M–14B parameters tested) scales better, improving up to 20% over the best SID configurations.
-
Enhancement-layer pattern: the LF AI & Data survey (August 2025) describes the prevailing non-generative architecture — LLMs generate semantic item/user embeddings and enrich features feeding conventional ranking stacks, i.e. content-based filtering re-implemented on foundation-model representations.
Sources:
-
https://research.atspotify.com/2025/9/semantic-ids-for-generative-search-and-recommendation