Collaborative Filtering

流媒体推荐系统经历了三代演进：第一代（2000年代）协同过滤、第二代（2010年代）深度学习、第三代（2020年代至今）大语言模型

AI can reorder menus by analyzing users' order history, stated preferences and even wearable health data, drawing on collaborative filtering or content-based filtering techniques from e-commerce and streaming platforms.

If popular items are not downweighted, similarity will be dominated by hit products, recommendations will converge, and collaborative filtering will degenerate into a bestseller list.

The core advantage of collaborative filtering is that it does not need to understand item content and recommends based solely on user behavior records.

If the same knowledge point is labeled with different spellings across different questions, recommendations based on collaborative filtering or vector retrieval will show severe bias.

This study essentially belongs to the content-based recommendation paradigm, and its conclusions cannot be directly extrapolated to collaborative filtering scenarios that require modeling user-item interactions; extending it to a complete recommendation system would also require solving the problems of how to encode user behavior sequences within the quantum circuit framework and how to connect with the classical recall/ranking pipeline.

Collaborative filtering relies on the user-item interaction matrix and makes recommendations under the assumption that 'similar users like similar items'; it typically captures finer-grained personalized signals but faces sparsity and cold-start problems.

Collaborative filtering faces the cold-start problem: recommendation performance drops significantly when new users or new dishes lack historical data.

Collaborative filtering is one of the most classic methods in recommendation systems, and its core idea is that similar users will have similar preferences.

The core principle of recommendation engines usually combines two methods, collaborative filtering and content-based recommendation; collaborative filtering relies on the accumulation of user behavior data, while content-based recommendation matches based on property features and customer preference profiles.

Collaborative filtering finds groups with similar tastes by analyzing large volumes of user historical behavior; its advantage is the discovery of latent preferences, but it suffers from a cold-start problem

The basic assumption of collaborative filtering is: if two users have highly similar past viewing preferences, content that one likes but the other has not watched is likely to also be liked by the other.

Memory-based methods of collaborative filtering directly compute the similarity between users or items (e.g., cosine similarity, Pearson correlation coefficient); they offer strong interpretability but degrade in performance when facing sparse matrices.

Model-based collaborative filtering uses matrix factorization to decompose a high-dimensional sparse rating matrix into two low-dimensional dense matrices representing user latent vectors and item latent vectors, and then predicts ratings via inner product.

Algorithmic recommendations rely on user behavior data (such as clicks, dwell time, and purchase history) for personalized matching, with the advantages of scalability and real-time performance.

History-data-based recommendation methods such as Collaborative Filtering tend to reinforce existing preferences, creating filter bubbles.

Collaborative Filtering is the foundational technology of traditional recommendation systems, based on the logic that you may also like what users with similar behavior to yours like.

Spotify's recommendation engine has long combined three types of technologies: collaborative filtering, natural language processing, and audio analysis.

Spotify's personalized recommendation system relies on multiple technologies including collaborative filtering and natural language processing, analyzing users' playback history, skip behavior, playlist-building habits, and comparisons with similar users.

Before the introduction of large language models, personalized recommendation systems long relied on two types of algorithms: collaborative filtering and content-based filtering.