How Xianyu Uses Real‑Time Data Analytics to Accelerate Operations

Background

Xianyu’s business growth brings a large number of activities and experiments. Traditional qualitative analysis can no longer quickly and accurately reveal the true state of the business, especially in user acquisition scenarios where deep data analysis is required.

Pain Points

When data problems arise, operations must request engineers for support. As Xianyu’s business becomes more complex, the number of requests (Request 1, Request 2, … Request N) grows, leading to long cycles that cannot meet business needs.

Solution Idea

Introduce Nanomirror to provide real‑time data analysis capabilities, democratizing data science. Build a “people‑goods‑scene” data system that encapsulates data collection, cleaning, and model development as APIs, offering dynamic drill‑down, intelligent facet analysis, and metric prediction.

Data Construction

The data system aggregates user, product, and strategy tags and metrics, forming the foundation for subsequent analysis. Additional data outside the tag library are obtained via a custom DSL from client‑side event tracking.

Features

Dynamic Drill‑Down

Steps:

Select a specific activity and metric.

Choose an interesting facet for secondary analysis and click the drill‑down button (e.g., sensitive user group).

View the analysis results for the combined facet (e.g., sensitive users + gender).

Intelligent Facet Analysis

Goal: Identify the facet combinations that produce the most significant metric improvements, helping operations quickly find reasonable sub‑populations for further intervention.

Content includes two parts:

Activity metric facet analysis – find the best facet combination for the activity itself.

AB‑bucket effect facet analysis – find the facet combination with the largest difference between experiment and control buckets.

Method combines correlation analysis, variance analysis, and decision‑tree logic, consisting of:

Correlation analysis to remove strongly correlated facets.

Clustering to discretize continuous facet values.

Information‑gain calculation to select the most informative facets recursively.

Pruning of nodes that do not meet sample‑size or gain thresholds.

Selection of the best facet based on effectiveness and minimum valid sample size.

Activity Metric Facet Analysis

Purpose: Find the facet combination that yields the highest activity effect.

Result includes the most significant facet combination, the metric value for each facet, and the minimum sample size required for reliable results.

Procedure:

Input dozens of facets and a single metric (e.g., gender, age, occupation vs. purchase rate, posting rate, retention).

Perform correlation analysis to drop redundant facets.

Apply clustering to convert continuous facet values into discrete categories.

Compute information‑gain ratio to identify the most informative facet and recurse.

Prune nodes that fall below population or gain thresholds.

Obtain the optimal facet combination and calculate the minimum effective sample size.

Example (data anonymized for security): In an activity reaching 1,000,000 users with an overall new‑purchase rate of 5 %, the facet “age 20‑30 & female” accounts for 200,000 users, meets the minimum sample size of 10,000, and shows a 10 % purchase rate.

AB‑Bucket Effect Facet Analysis

Goal: Identify facet combinations that experience the greatest positive or negative impact from the experiment.

Steps (similar to activity analysis):

Input facets and metric.

Remove strongly correlated facets via correlation analysis.

Cluster continuous facet values.

Select reasonable facet combinations, compute minimum effective sample size, and discard those below thresholds.

Extract the facet combinations with the largest and smallest metric differences.

Result example (anonymized): The AB‑bucket analysis reveals three facets with the highest uplift in the “new‑purchase” metric, allowing operators to increase spend on positively impacted groups (e.g., females aged 30‑40) and reduce spend on negatively impacted groups.

Metric Prediction

Beyond post‑experiment analysis, Nanomirror can predict key metrics before an activity launches, shortening iteration cycles. Operators input an activity ID and analysis window, and the system returns predicted outcomes.

Example: The platform accurately forecasted the purchase rate for the “222” promotion, enabling timely adjustments to creative assets.

Outlook

Future work aims to build a knowledge base and simulation engine for Xianyu, further reducing operational costs and shortening activity iteration periods.