Comprehensive Overview of Ranking Models in Recommendation Systems

Introduction – Ranking (精排) is a crucial module in recommendation algorithms. It is usually model‑driven and involves three core components: samples, features, and models. This article provides a detailed exposition to help developers.

1. Overall Architecture

Ranking is a key part of the recommendation pipeline.

2. Samples

Samples are the "food" for models. In a CTR task, exposures with clicks are positive samples, exposures without clicks are negative samples. Main issues include:

Imbalanced positive/negative samples : e.g., a 5% click‑through rate yields a 1:20 positive‑negative ratio. Solutions: negative sampling, focal loss, etc.

Different activity levels : long‑tail users have few samples, high‑frequency users dominate, leading to bias. Mitigation: down‑sample active users or enforce uniform sample counts.

Sample confidence : an unclicked exposure does not always mean the user dislikes the item; sometimes the user has no intent. Strategies include filtering all‑negative users.

Invalid traffic : bot or crawler traffic should be filtered at the recommendation‑engine side.

User intent saturation : Meituan uses "skip‑above" sampling to drop negatives after the last click.

3. Features

Feature engineering remains essential despite the rise of deep models. Typical feature groups:

Context features : weekday, hour, network type, OS, client version, etc.

User features (static): user_id, gender, age, city, occupation, income, student status, marital status, registration time, VIP flag, new‑user flag.

User statistical features : recent 30/14/7‑day PV, VV, CTR, completion rate, average view time (both absolute and relative). Beware of data leakage.

User behavior sequence features : short‑term click sequences, long‑term purchase sequences, positive‑feedback vs. negative‑feedback sequences. Sequence length is a bottleneck for Transformer‑based models.

Item features (static): item_id, author_id, category_id, publish time, resolution, duration, tags.

Item statistical features : recent PV, VV, CTR, completion rate, average view time (same leakage caution).

Cross features : item‑user cross statistics (e.g., CTR per gender/age group).

Feature processing methods:

Discrete values : direct embedding; watch out for high‑dimensional sparse IDs.

Continuous values : (a) concatenate with embeddings (low generalization) or (b) equal‑frequency binning then discretize (better generalization).

Multi‑value features (e.g., behavior sequences): mean‑pooling, sum‑pooling, attention‑pooling (DIN), RNN/GRU‑based sequence modeling, Transformer‑based modeling.

4. Models

Evolution from linear to deep models:

Collaborative Filtering (CF) : userCF, itemCF, matrix factorization (MF). MF reduces sparsity by factorizing the user‑item interaction matrix.

Logistic Regression (LR) family : LR, FM, FFM, LS‑PLM. LR can incorporate rich user/item features but lacks feature crossing (Simpson’s paradox).

Deep Neural Network (DNN) family : DeepCrossing, FNN, PNN – embed sparse IDs, stack embeddings with numeric features, feed into multi‑layer DNN with residual connections.

Wide & Deep (WDL) family : combines LR‑style wide part (memory) with DNN deep part (generalization). Variants include DeepFM, xDeepFM, NFM, etc.

Attention & Sequence models : DIN (attention pooling), DIEN (interest evolution with GRU + target attention), MIMN, SIM – address multi‑value feature modeling and user interest dynamics.

5. Ranking Optimization

Key optimization dimensions:

Memory vs. Generalization : high‑frequency items rely on memorization, low‑frequency items need generalization (addressed by WDL).

Feature & Cross Engineering : manual cross features (user‑item statistics) vs. automatic cross (FM, DeepFM, xDeepFM).

Embedding of High‑Dimensional Sparse IDs : crucial but challenging to converge for long‑tail IDs.

Personalized Behavior Modeling : sequence‑based user interest modeling (attention, RNN, Transformer) with efficiency tricks for long sequences.

Author

Xie Yangyi – Tencent Application Algorithm Researcher, responsible for video recommendation algorithms, with extensive experience in NLP and search/recommendation.

Recommended Reading

揭秘一致性Hash算法应用！

手把手教你从0开始实现C++协程！

保姆级教程！Golang微服务简洁架构实战

阅见深我，读享生活，TVP读书分享会带你解锁新知！