How We Built a Transparent, Explainable Traffic Allocation System for E‑Commerce

Background

In the Yanxuan e‑commerce platform, traffic is distributed across multiple modules (search, recommendation, placement, etc.). Deep‑learning models improve performance but are opaque, making it difficult for business teams to intervene, which can cause bias and inefficiency. To obtain a Pareto‑optimal balance between traffic efficiency and fairness, a transparent, explainable traffic‑decision system was built.

Process

Direction

The system integrates marketing and traffic‑allocation logic instead of relying solely on a model‑driven loop. Key principles are:

Divide products into logical pools (new, potential, seasonal, supplement) to make traffic flow more reasonable.

Apply multi‑dimensional, multi‑scenario, user‑segmented support policies that match traffic supply with user‑demand rhythm and activity cadence.

Balance absolute and relative demand, stage‑specific growth, and fairness through a PID‑like control.

Traditional interest‑driven distribution (product representation, click‑through estimation, multi‑task learning, cold‑start, scattering) is extended with business‑driven capabilities such as pool division, traffic allocation, and joint support.

Methods

Selection Capability

Product evaluation goes beyond simple SABC sales grading. Two extra dimensions are introduced:

Perspective dimension : metrics are collected from both business‑unit (BU) and user viewpoints.

Spatio‑temporal dimension : dozens of metrics (clicks, dwell time, search, conversion, rating, return, repurchase, etc.) are normalized and a wave‑corrected trend metric is added to capture temporal dynamics.

Products are grouped into pools—new‑product, potential, seasonal, supplement—so that downstream traffic‑support policies can be applied per pool.

Targeting Capability

Three targeting logics are defined:

Attribute‑based targeting : e.g., gender‑specific recommendations.

Interest targeting : shallow interaction, explicit demand, and candidate demand, covering ~20 user‑need types.

Seasonal trigger targeting : regional search volume of seasonal keywords is used to infer demand; when a region’s seasonal demand exceeds a threshold, flow support is activated for seasonal products in that region.

PV Value Estimation

PV value quantifies a product’s traffic‑handling capacity. Real‑time PV estimates are combined with user‑interest scores to decide whether to use insertion‑type or position‑adjustment‑type control.

Allocation Algorithm

The algorithm solves two sub‑problems:

How much to allocate : allocation considers product growth stage, absolute vs. relative demand, and a PID‑like controller that balances absolute volume, relative gap, stage error, and cumulative error.

How to allocate : two modes are used:

Fixed‑position mode (short‑term burst): traffic is inserted directly into flexible slots using filtering, ranking, and probabilistic sampling. Suitable for new and trending products.

Dynamic‑position mode (endurance): allocation is treated as re‑ranking; position shifts are computed from real‑time feedback and applied with probabilistic sampling and density checks. Suitable for potential products that need gradual support.

Other Capabilities

Data pipelines, engineering infrastructure, and visualization tools connect all stages. Real‑time A/B‑test data, multi‑level KV storage, conflict handling, and dashboards enable seamless collaboration between engineers and operators.

Conclusion

The project follows a design‑thinking workflow: analyze user, operation, and model constraints; break the pure‑model iteration loop; prototype and iterate traffic‑control strategies per module; then scale to multi‑module coordinated control, continuously refining based on feedback.