How Intelligent Slice Analysis Transforms Xianyu’s Operational Decision‑Making

Background

In the era of big data, Xianyu’s operations lacked deep data‑driven analysis capabilities, leading to repeated experiments and missed insights into sub‑segment performance.

Pain Points

Historical data is not accumulated, causing a cycle where new operators repeat similar experiments.

Insufficient granular analysis of activities, e.g., overall metrics look flat while specific sub‑segments (age 20‑30, female) show strong effects.

Intelligent Slice Analysis Overview

Nanomirror’s decision‑analysis module offers “intelligent slice analysis” that automatically discovers the most significant user‑segment combinations for a given metric, enabling operators to focus on high‑impact sub‑populations.

Two Core Components

Activity‑Metric Slice Analysis – finds the best segment combinations for the activity itself.

AB‑Bucket Effect Slice Analysis – identifies segment combinations where the experimental bucket differs most from the control bucket.

Activity‑Metric Slice Analysis

The goal is to locate the segment combination that yields the highest activity effect.

Collect results: obtain the most significant segment combinations and their metric values.

Determine minimum sample size: flag results with insufficient data as uncertain.

Steps to generate these results:

Data analysis: input dozens of slices (e.g., gender, age, occupation) and a single metric (e.g., purchase rate, retention).

Correlation analysis: evaluate relationships among slices and metrics, removing strongly correlated slices (e.g., 7‑day vs. 14‑day purchase counts).

Clustering: convert continuous slice values into discrete categories via clustering algorithms.

Information‑gain calculation: compute the information‑gain ratio for each slice, recursively selecting the highest‑gain slices.

Pruning: discard nodes with sample sizes or gain ratios below thresholds.

Best slice extraction: assess slice effectiveness and calculate the minimum effective sample size.

Example result (illustrative, non‑real data): among 1,000,000 participants, the overall “new purchase today” rate is 5 %. The segment age 20‑30 & female accounts for 200,000 users, meets the minimum sample size of 10,000, and shows a 10 % purchase rate, indicating a strong positive effect.

AB‑Bucket Effect Slice Analysis

This analysis seeks the segment combinations where the experiment’s impact is maximal (positive or negative), allowing operators to amplify exposure to positively affected groups and reduce it for negatively affected ones.

Steps (mirroring the activity‑metric analysis):

Data analysis: input slices and a metric.

Correlation analysis: remove redundant slices.

Clustering: discretize continuous slice values.

Select reasonable slice combinations: compute minimum effective sample size and filter out low‑population or low‑confidence slices.

Best slice extraction: identify the combinations with the greatest metric differences between experiment and control buckets.

Result example: three buckets show the largest lift in the “new purchase today” metric, each with specific segment definitions (e.g., sensitive users = 1, 30‑day interaction days = X) and corresponding confidence‑adjusted improvements of ~0.8 %.

Conclusions

Nanomirror can now analyze Xianyu’s existing activities (e.g., 222, red‑packet, guide‑sale campaigns) by simply entering an activity ID and time window, producing actionable segment insights.

Outlook

Future work aims to simulate operational outcomes using existing knowledge, reducing operational costs and shortening activity iteration cycles.