iQIYI Data Quality Monitoring: Exploration and Practice

On September 26, iQIYI's technology product team hosted the 19th "iTech Salon" with the theme "Data Governance Exploration and Application". Senior experts from Kuaishou, Meituan, and Kuakan shared insights.

Researcher Peng Tao presented "Exploration and Practice of iQIYI Data Quality Monitoring", focusing on the rule‑engine module of the data‑governance platform, its current challenges, goals, anomaly‑detection methods, and future enhancements.

Problem and Goal: Why monitor data quality?

Data quality monitoring is likened to epidemic prevention: early detection (like nucleic‑acid testing) and tracing (like source tracing) help identify and control issues.

Data anomalies stem from three main factors:

Product factors: e.g., app releases that change Pingback delivery strategies.

Operation & external factors: channel operations, content flow, artificial traffic, partner influences.

Technical issues: missing data, calculation logic errors.

iQIYI addresses these through three monitoring layers:

Pingback layer: the source of all reports; improves delivery quality at the source.

Data middle layer: adds necessary checks to prevent abnormal data from propagating downstream.

Business report layer: monitors core business metrics (PV, UV, field null‑rate, etc.) to ensure health of downstream analytics.

Monitoring dimensions and indicators

Pingback monitoring is split into three dimensions:

Business dimension: specific products/endpoints (e.g., iQIYI Android, iPhone).

Event‑type dimension: user actions such as start, play, click.

Time dimension: 5‑minute, hourly, daily granularity.

Metrics are standardized (log PV/UV, field null‑rate, mean values, enum distributions) to enable automated monitoring.

Anomaly detection methods

Various detection techniques are employed:

Threshold method: simple upper‑limit rules (e.g., CTR > 98%).

Box/Gaussian detection: statistical approach using ±3σ from a 30‑day window; dynamic but may cause over‑fitting.

Correlation detection: compares two related metrics (e.g., conversion = B/A); flags low correlation (<0.8) as abnormal.

Prophet time‑series forecasting: predicts values with confidence intervals; anomalies are points outside the interval.

Each method has pros and cons; for instance, correlation detection is good for cross‑metric comparison but cannot pinpoint which metric is faulty.

Future plans

iQIYI aims to develop:

Intelligent detection: automatically generate monitoring rules from historical trends (except for correlation metrics that require manual setup).

Intelligent attribution: after an anomaly is detected, drill down dimensions to identify the most impactful factors.

The intelligent attribution architecture includes dimension‑drill management, a data‑graph for lineage, expert‑advice repository, and an attribution engine.

Overall, the talk provided a comprehensive view of iQIYI's data quality monitoring framework, the challenges faced, and the roadmap toward smarter detection and attribution.