Machine Learning-Based Anomaly Detection for Core Business Metrics

This paper presents a machine learning-based solution for detecting anomalies in core business metrics, specifically focusing on daily active users (DAU) for a cloud music platform. The project addresses limitations in traditional threshold-based monitoring systems that struggle with dynamic business changes and varying seasonal patterns.

The research evaluates multiple anomaly detection algorithms including 3-sigma, Holt-Winters, and XGBoost, comparing their performance using metrics like recall, precision, accuracy, and F1-score. Results show XGBoost with multiple features achieves superior performance with 89% recall and 81% precision, significantly outperforming traditional methods.

The solution integrates machine learning models with existing data quality monitoring systems through a containerized, configurable architecture. The approach combines rule-based detection with model predictions using a fusion strategy to improve robustness. The system demonstrates 74% improvement in recall, 40% in precision, and 20% in accuracy compared to baseline methods.

Future work includes model iteration for reducing false positives, expanding applicability to more business scenarios, and developing a staged approach for handling new metrics with limited historical data. The project emphasizes the evolution from digital to intelligent data systems and the importance of open-source contributions to data intelligence.