Deconstructing E‑commerce Recommendation Systems: Architecture, Challenges, and Strategies

The talk, presented by Yao Kaifei (CEO of Guanbian Technology) and edited by Chen Jiahui, introduces the importance of recommendation systems in the mobile internet era, where massive user data and big‑data technologies enable personalized product suggestions that reduce information overload for users and drive precise marketing for businesses.

1. Work Flow Overview – A complete e‑commerce process includes browsing, clicking, adding to cart, purchasing, and long‑term repurchase, with the recommendation subsystem matching users to items at each stage. The system’s value lies in efficiently connecting users with relevant products.

2. Challenges – Multiple scenarios and objectives require balancing short‑term behaviors (clicks, conversions) with long‑term goals (retention, repeat purchase). Data must be collected across user, product, and merchant dimensions, forming a full‑chain data loop.

3. Key Modeling Factors – User journeys are modeled from initial click through conversion and eventual repurchase. Modeling focuses on click‑through rate (CTR), conversion rate (CVR), price, and other factors, often combined as weighted formulas (e.g., CTR^α × CVR^β × Price^γ) to estimate GMV.

4. System Architecture – The recommendation pipeline consists of recall (real‑time behavior, historical behavior, profile‑based, hot‑item/trend) followed by filtering (exposure, purchase, gender, etc.), coarse ranking (CTR/CVR estimation, diversity, novelty), and fine ranking (business rules, diversity control, reinforcement learning for parameter tuning).

5. Recall Strategies – Real‑time behavior recall captures immediate user actions; historical recall leverages offline behavior; profile recall uses static user attributes; hot‑item recall surfaces trending products. Effective recall provides high‑quality candidates for downstream ranking.

6. Ranking Strategies – Ranking incorporates numerous factors (CTR, Cvr, price, match scores, time decay) and may differ across platforms (e.g., short‑video vs. e‑commerce). Multi‑objective optimization balances short‑term conversion with long‑term ecosystem health.

7. Model Evolution – Models have progressed from daily batch updates to hourly and near‑real‑time systems, now employing deep learning architectures (embedding, LSTM, attention, Transformer) to capture sequential user behavior and improve interest prediction.

8. Feature Construction – Features include user and item profiles, behavior sequences, and matching features (e.g., DIN, DIEN). Both static and dynamic features are integrated into ranking models.

9. Evaluation – Recall is evaluated by hit rate, precision, coverage, and F1‑score, while ranking impact is measured through downstream metrics such as GMV, conversion, and user retention. End‑to‑end evaluation considers the influence of upstream modules on downstream performance.

10. System & Global Ecosystem – Building a recommendation system requires aligning short‑term revenue goals with long‑term ecosystem health, considering merchant tiering, product lifecycle, intelligent traffic allocation, and multi‑objective optimization across users, merchants, and the platform.

11. Practical Discussions – Topics include user segmentation, same‑store behavior analysis, cart recommendation strategies, and the importance of hypothesis‑driven experimentation, data‑driven iteration, and healthy metric design.

The presentation concludes that effective recommendation systems are not necessarily the most complex algorithms but follow a disciplined workflow of hypothesis, analysis, strategy, and measurement, continuously iterating based on real‑time data and business objectives.