Quago builds ML behavioral analytics on Databricks to detect fraud and cheating in online games

Overview

Quago analyzes over 200 device-sensor data points per player swipe using machine learning models built on Databricks Delta Lake and Spark to detect cheating, churn and user-acquisition fraud in real time for gaming companies. Migrating from a vanilla Parquet data lake to Databricks let Quago run SQL queries 4x faster, reduced customers' server costs by an estimated 25%, and is projected to save gaming companies tens of millions of dollars in in-app purchase fraud losses.

The challenge

Quago originally used a vanilla Parquet lake that was costly to scale due to performance, storage efficiency and query optimization challenges, and struggled with cost-effective scaling and the resource-intensive nature of managing infrastructure for both performance and availability with a small engineering team.

The solution

Quago migrated its data pipeline into Databricks Delta Lake, using Databricks SQL and Apache Spark on top of Delta Lake to build, retrain and launch ML fraud, cheating and churn-detection models, updating models per customer and app as often as three times a week.

Reported business value

Quago achieved 4x faster SQL queries compared to its previous vanilla Parquet lake, reduced customers' server costs by up to 25%, and estimates its fraud mitigation will reduce customer in-app purchase losses by tens of millions of dollars, while keeping a small, efficient engineering team.