Key Characteristics and Practical Improvements of Recommendation Technologies

Recommendation systems aim to predict a user's preference probability for items (e.g., news) when no explicit behavior exists, then rank items by this probability. The process typically leverages user profile features, content features, contextual scene features, and various user interaction behaviors to score candidate items.

Two classic collaborative filtering approaches are UserCF and ItemCF. ItemCF suits personalized scenarios (e.g., products, movies, short videos) by focusing on the current user's recent actions, while UserCF is more group‑oriented, favoring popular items when many similar users are considered. ItemCF provides weaker interest expansion than UserCF and solves user cold‑start, whereas UserCF addresses item cold‑start.

From a mathematical perspective, collaborative filtering works because rows (or columns) of the interaction matrix that are partially similar tend to be linearly related, implying a low‑rank structure. Exploiting this redundancy allows matrix compression without significant information loss, which is a core challenge in recommendation research.

Matrix factorization demonstrates this low‑rank property by decomposing a large sparse matrix into the product of two smaller matrices. The loss function E combines an empirical loss term and a structural risk term, guiding the factorization.

Although CF and matrix factorization are closely related, industrial practice often faces issues such as extreme sparsity, over‑fitting, and high computational cost. Consequently, many variants and side‑information‑enhanced methods have emerged.

For item2vec, the common algorithms are ALS and Word2Vec. Word2Vec, which learns sequential embeddings of users and items, often yields a 30%+ click‑through improvement over ALS in news feed scenarios because it captures diverse local behaviors and employs high‑frequency‑biased negative sampling.

Negative sampling differences: ALS performs global negative sampling, while Word2Vec focuses on high‑frequency samples, reducing the tendency for the model to over‑emphasize head items.

In CTR prediction, a triple of (previous click, next impression, label) is used to estimate the click probability between two item IDs, directly optimizing the click‑through objective. However, this approach struggles with cold‑start items due to lack of historical signals.

To address cold‑start, side information (e.g., item ID, keyword, category, author) can replace the raw item ID in a skip‑gram model, turning Word2Vec into a multi‑class classifier that learns richer item representations; similar side information can be incorporated on the user side.

Factorization Machines (FM) extend logistic regression by adding pairwise feature interactions via latent vectors, effectively handling data sparsity. FM can be used for both CTR ranking and recall tasks because its interaction parameters are shared across feature pairs.

In practice, the best performance is achieved by ensembling multiple models (e.g., various CTR estimators) and feeding their dense features into a GBDT, which captures complex interactions.

Real‑world e‑commerce systems often need to balance CTR and conversion rate (CVR). Separate high‑accuracy CTR and CVR models are built, and a weighted ranking formula combines their scores; hyper‑parameters for this combination are frequently tuned via reinforcement learning, yielding several percentage points of conversion uplift.

Greedy ranking models tend to over‑expose high‑PV, high‑click items, neglecting long‑tail high‑quality content. A mature solution is to apply personalized Upper‑Confidence‑Bound (UCB) algorithms, which balance exploitation and exploration to surface hidden valuable items.

In summary, recommendation technology differs from search: recommendation seeks novelty and personalization based on abundant user signals, while search focuses on precision from explicit queries. Deploying these techniques successfully requires careful feature engineering, iterative experimentation, and attention to engineering details.