From the Pre‑Recommendation Era to the Bronze Age: Evolution of Recommendation Systems and Mitigating the Matthew Effect

The piece begins by defining the "pre‑recommendation era" as a time when recommendation functions were simple, global, and lacked personalization, relying on manual product selection and offline recall‑ranking logic.

It then details three main characteristics of this stage: (1) simple global recommendation without personalization; (2) offline‑centric recall and ranking with lightweight service logic; and (3) reliance on manually crafted or partially machine‑learning‑assisted strategies.

1. Manual Sorting and Operations

Early product recommendation depended on operators manually configuring item lists based on business knowledge, adjusting rankings by SKU count, region, gender, and other demographic factors. While feasible for small catalogs, this approach becomes untenable as the number of SKUs grows into the tens of thousands.

1.2 Real‑time Hotspots

Human intervention remains necessary for sudden events (e.g., World Cup, Olympics) where timely hot‑topic items must be injected into recommendation lists.

2. Natural Ranking

Natural ranking emphasizes three principles—hot, fast, and complete—prioritizing popularity, recency, and eventual personalization. Simple hot‑ranking can be generated from multi‑dimensional popularity metrics such as click‑through or purchase leaderboards.

2.2 Example

In B2C e‑commerce, ranking may combine factors like sales volume, inventory depth, novelty, and price, often using weighted formulas (e.g., a × (1‑b) × (0.5c + 0.1d + 0.03e + 0.2f)).

3. Machine‑Learning‑Based Scoring

Data points (exposures, clicks, adds‑to‑cart, purchases) are collected to train models that predict item conversion probabilities, allowing fine‑grained ranking while still supporting manual weight adjustments.

Mitigating the Matthew Effect

The article discusses the "Matthew effect"—the rich‑get‑richer phenomenon— and proposes remedies such as random insertion of new items, periodic down‑weighting of top‑ranked items, and similarity‑based score inheritance for cold‑start products.

Bronze Age: Association and Personalization

Transitioning to the "bronze age," systems incorporate association (item‑item similarity) and personalization (user‑item matching) to address information overload and long‑tail exposure, leveraging large‑scale data, user behavior modeling, and multi‑objective ranking.

Personalization Workflow

Typical steps include i2i data generation (behavior weighting, collaborative filtering), candidate recall (similar items based on recent actions), model‑based scoring (CTR prediction), and diversification (ensuring category variety).

System Architecture

A modular pipeline—recall, filter, ranking, re‑ranking—supported by log collection, offline/near‑real‑time computation, and micro‑service deployment ensures scalability and cost‑effectiveness.

Recall Strategies

Recall sources span context‑related (time, location, scenario), interest‑related (user profile, long‑/short‑term interests), behavior‑related (collaborative filtering), and hot‑/supplementary lists to guarantee coverage.

Conclusion

Modern recommendation systems rely on multi‑route recall, diversified ranking, and continuous model iteration to balance relevance, diversity, freshness, and fairness, forming the foundation for subsequent personalization and reinforcement‑learning advancements.