Diagnostic Analytics in Meituan Food Delivery: Methods and Case Studies

Guest Speaker: Wang Qing, Data Analysis Expert at Meituan

Editor: Liu Xue, ByteDance

Platform: DataFunTalk

Introduction: Why do metrics rise, fall, or stay flat? In a competitive market, quickly and accurately locating the cause of metric changes is increasingly important. This article shares practical experience of diagnostic analysis in the food delivery business.

The discussion is organized into four parts:

Why perform diagnostic analysis?

How to conduct diagnostic analysis?

Intelligent diagnostic case study

Tips: common pitfalls in diagnostic analysis

01 Why Perform Diagnostic Analysis

Diagnostic analysis is valuable for two main reasons: finding the root cause of a problem and identifying highlights to promote.

Value 1 – Find the Root Cause Identifying the reason behind a problem provides direction for strategy and solution. Examples include Google’s 45‑minute outage in Dec 2020 (≈$1.7 M loss) and a recent Douyin login issue affecting user experience and ad revenue.

Not every issue needs a root‑cause analysis; some are random (e.g., a punctured tire) or have limited business value (e.g., discovering a segment of users who like spicy food).

Value 2 – Find Highlights Once a highlight is discovered, it can be promoted. Recommendation systems often excel at leveraging such highlights.

The focus of this talk is on Value 1 – finding the root cause.

02 How to Perform Diagnostic Analysis

1. What is Diagnostic Analysis? Gartner defines it as “Why did it happen?” In practice, diagnostic analysis breaks down a problem using data and compares components to uncover the cause.

2. What is a Problem? A problem is the gap between the current state and the expected state. Problems can be classified into three types:

Occurrence‑type: e.g., why did order volume drop 10% compared to yesterday?

Potential‑type: e.g., is weather affecting staffing needs?

Ideal‑type: e.g., how to achieve a Q3 OKR target increase from 10% to 12%?

We will mainly discuss occurrence‑type and potential‑type problems.

3. How to Conduct Diagnostic Analysis

A widely applicable method is “Logical Tree + Hypothesis‑Driven”. The logical tree structures the metric’s internal composition, while hypothesis‑driven defines the analytical perspective.

The process:

Clarify the gap between current and expected values.

Decompose the problem (not just the metric) into two trees: one based on expected proportion, the other on actual proportion.

Compare the two trees to identify differences and infer causes.

Example: Order volume dropped 10% (60 orders). By hypothesizing city‑level impact, we find Beijing contributed 50 of the lost orders, and within Beijing, Chaoyang district contributed 40. Further analysis shows weather (rain vs. sunny) caused a 42‑order drop, pinpointing weather change as the main cause.

03 Intelligent Diagnostic Case Study

Manual diagnostic analysis is labor‑intensive; therefore, algorithmic and modeling approaches are needed to automate the process.

Project 1: Weather Index

Background & Goal Food delivery heavily depends on offline conditions; weather influences DAU (rain increases, pleasant spring decreases). The goal is to build a weather‑impact index for DAU across cities to aid variance analysis.

Solution

1. Transform the variance problem into a composition problem by splitting DAU into three parts: weather‑driven DAU, other‑factor DAU, and natural growth DAU (trend + seasonality).

2. Identify influencing factors: weather (temperature, humidity, phenomena, secondary disasters) and other factors (business strategy, holidays).

3. Build the index through five steps: data acquisition, preprocessing, target construction, modeling (XGBoost), and aggregation. City‑level modeling is used because weather impact varies by region; clustering similar climate cities mitigates sample scarcity.

Project 2: Intelligent Diagnostic System

Background & Goal Business needs rapid detection of metric anomalies and root‑cause location. Manual analysis is costly and requires high expertise. The goal is to build an intelligent system that automatically identifies anomalies and diagnoses their causes.

The system consists of two modules: anomaly detection and anomaly diagnosis.

Solution Design

1. Anomaly detection determines whether a metric deviates from expected behavior (e.g., comparing week‑over‑week order volume). 2. Anomaly diagnosis answers “why” by performing both qualitative (correlation, event analysis) and quantitative (contribution, composition analysis) investigations. The system provides a logical‑tree builder and hypothesis‑driven interface for users to specify analysis perspectives.

The diagnostic algorithm identifies abnormal dimensions (similar to Gini coefficient) and then pinpoints the specific dimension values that contribute most to the anomaly.

04 Common Pitfalls in Diagnostic Analysis

Failures often stem from unclear problem definition or staying within a “comfort zone”. A clear problem requires a well‑defined current state, expected state, and the gap between them. Analysts must avoid limiting their view to a narrow set of dimensions, which can lead to missed or incorrect diagnoses.

In summary, a systematic, logical‑tree‑based, hypothesis‑driven approach, combined with automated tools, increases the likelihood of correctly identifying root causes and improving decision‑making.

Thank you for listening.