Evolution of Recommendation System Architecture: From Hand‑Crafted Rules to the Large‑Model Era

1. Algorithm Iteration and System Engineering Upgrade

Current mainstream recommendation algorithms are deep‑learning models such as DeepFM, DCN, DIN, and Transformers, with some large teams adding large models for fine‑ranking. The author stresses that algorithm upgrades must be accompanied by corresponding engineering system upgrades; only when existing algorithms hit business bottlenecks should a new algorithm be adopted.

2. Position of Algorithms in the Recommendation Pipeline

The online recommendation flow is simplified as:

Request → Recall → Coarse‑ranking → Fine‑ranking → Re‑ranking → Result

. The five generations (rules, LR, GBDT, deep learning, large models) cover all these stages and coexist rather than replace each other.

3. First Generation – Manual Rules

3.1 Concrete Implementation

Hotness ranking – prioritize content with high clicks, views, interactions.

Dimension filtering – filter by channel, category, region, publish time.

Manual weight intervention – boost high‑quality or hot content, suppress low‑quality or violating content.

Basic tag matching – match static user tags (age, gender, region, device) to content.

Manual weight fusion – combine multiple dimensions with a fixed formula, e.g.,

Score = ClickRate×0.4 + Freshness×0.3 + CategoryMatch×0.2 + RegionMatch×0.1

.

3.2 Scenarios and Value

Rules work well for early‑stage cold‑start where user and content volumes are low. Advantages include low development cost, fast iteration, efficient online execution, strong controllability, and high interpretability.

3.3 Limitations

Rule explosion – more scenarios lead to ever‑growing rule sets.

Rule conflicts – overlapping rules cause unstable results.

Inability to personalize per user (“one‑size‑fits‑all”).

Sub‑optimal static weights.

Rising maintenance cost.

Thus, rules are suitable for 0→1 growth but cannot support 1→100 personalization.

4. Second Generation – LR Linear Model

4.1 Why Move from Rules to LR

As user numbers and feedback increase, static rule weights become mismatched. LR learns weights from data, automatically adapting to changing user preferences.

4.2 Value of LR

Data‑driven weight learning replaces manual tuning.

Adapts to different time slots, users, and emerging hotspots.

Training is fast, inference is lightweight, and the model remains highly interpretable.

4.3 Engineering Upgrades Triggered by LR

Introducing LR requires a data‑processing pipeline (storage, cleaning, sample generation), offline feature production, online inference modules, A/B testing framework, and model version management.

5. Third Generation – GBDT Tree Model

5.1 Linear Model Shortcomings

LR cannot capture feature interactions or non‑linear relationships, relying heavily on manual feature engineering.

5.2 Breakthrough of GBDT

GBDT automatically learns non‑linear feature combinations, reducing manual effort and improving ranking precision.

5.3 Engineering Costs of GBDT

Higher training and inference latency (serial tree construction, many tree traversals).

Increased memory and storage due to multiple trees.

Complex deployment (leaf encoding, one‑hot conversion, feature alignment).

More extensive hyper‑parameter search and sensitivity to data distribution.

Dedicated inference engines (e.g., XGBoost, LightGBM) increase scaling difficulty.

6. Fourth Generation – Deep Learning

6.1 Bottlenecks of Traditional Models

Massive user/content scale.

Sparse ID features.

Short‑term vs. long‑term interest divergence.

Emergence of multimodal content.

Increasingly complex scenario logic.

6.2 Value of Deep Learning

Embedding‑based efficient handling of high‑dimensional sparse IDs.

Automatic high‑order feature combination.

End‑to‑end modeling of images, audio, video via CNN/RNN/Transformer.

Transformer‑based short‑term interest modeling.

Multi‑objective joint optimization.

Better generalization through dense representations.

Pre‑training and transfer learning reduce cold‑start cost.

6.3 Engineering Upgrades for Deep Learning

Compute upgrade – distributed GPU clusters for training.

Feature platform – vector, sequence, and multimodal feature governance.

Inference acceleration – model compression, latency optimization.

Enhanced monitoring, versioning, and debugging pipelines.

7. Fifth Generation – Large‑Model Era

7.1 Limitations of Conventional Deep Models

They lack deep semantic understanding, logical reasoning, and strong zero‑shot generalization.

7.2 Capabilities Brought by Large Models

True content comprehension (semantic and emotional).

Implicit intent inference (e.g., “user viewed basketball shoes → likely needs sports socks”).

Stronger generalization for new or niche items.

Prompt‑driven flexible iteration.

7.3 Practical Deployment Choices

Due to extreme compute cost and hallucination risk, large models are typically used for re‑ranking or offline feature generation rather than replacing the entire pipeline. A common pattern is “deep learning for primary ranking + large model for final re‑ranking or feature enrichment”.

8. Panorama Comparison of the Five Generations

The author summarizes each generation across four dimensions—core capability, core problem solved, engineering complexity, and suitable business stage—showing a progression from low‑cost, rule‑driven cold‑start solutions to high‑cost, high‑capability large‑model enhancements.

9. Architecture Selection – Business Fit Over Technology Hype

Choosing an algorithm is a trade‑off among business impact, engineering capacity, and cost. Older techniques (rules, LR, GBDT) remain valuable as fallbacks or for specific scenarios, while deep learning is mainstream for mature, high‑traffic services, and large models are reserved for premium, resource‑rich deployments.

10. Summary and Next Preview

The article recaps the evolution path: manual rules → LR (data‑driven weight learning) → GBDT (non‑linear feature combination) → deep learning (dense representation and multi‑modal modeling) → large models (semantic understanding and reasoning). It emphasizes that model upgrades drive parallel upgrades in data pipelines, feature systems, and infrastructure. The next article will focus on feature engineering.