AppsFlyer Builds a Predictive Analytics Solution for iOS 14+ Using Amazon SageMaker
The advertising industry has been upended by new standards for data privacy, cookies, and usage of ad identifiers. To improve the measurement of marketing campaigns in this privacy-centric landscape, marketing measurement company AppsFlyer used Amazon Web Services (AWS) to deliver PredictSK, a predictive analytics solution that uses machine learning (ML) to accurately predict mobile user lifetime value (LTV) of iOS SKAdNetwork campaigns based on anonymous data to prevent specific user identification.
PredictSK uses ML and serverless services from AWS, including Amazon SageMaker, which helps data scientists and developers prepare, build, train, and deploy high-quality ML models quickly. The product uses predictive modeling to avoid tracking users while providing higher accuracy on campaign performance, producing insights based on the first 24–48 hours of user interaction. The solution also protects user privacy in compliance with Apple’s iOS 14 privacy changes for advertising.
The serverless architecture on AWS reduces development time and maintenance. We can start with a small scale but then increase the scale with the confidence that everything will still work.”
Senior Software Developer, AppsFlyer
Solving an Industry Challenge Rapidly on AWS
AppsFlyer has over 12,000 customers globally across ecommerce, financial services, gaming, and more, and its services are used in over 89,000 mobile apps. PredictSK uses ML to predict the LTV of mobile app users on IOS 14+ in a marketing campaign based on SKAdNetwork signals. The company began building its initial solution in the fourth quarter of 2019. “PredictSK significantly optimizes savings and campaigns for marketers,” says Michel Hayet, senior product marketing manager for predictive analytics at AppsFlyer. “It gives our customers a lot of knowledge about their campaign performance much sooner than the traditional LTV evaluation cycle, which can take weeks or months to know if a campaign is working well.”
Starting in 2018, data privacy regulations in the European Union and the United States have set new guidelines that restrict the types of personal information that can be collected, shared, and used for marketing. Major browsers have deprecated third-party cookies used for tracking, targeting, and measuring data. In June 2021, Apple rolled out iOS 14, which features SKAdNetwork, a framework for privacy-preserving mobile install attribution, and App Tracking Transparency, a framework that prevents consumers’ device identifiers from being used in ad tracking and targeting without their consent. “With iOS 14, we only have about 24 hours to look at each user’s behavior and decide if the user will be valuable and whether to invest in the media source that provided the user,” says Elena Levi, product team lead for PredictSK. “But with predictive analytics, all we need is 24 hours to get the long-term insights required.”
AppsFlyer used AWS to adapt its predictive analytics solution to meet new industry requirements. The solution went from an initial idea to a working product in 6–8 months. On AWS, AppsFlyer cut time to production by 66 percent with the same number of staff, and in June 2021, the company prepared to onboard the first customer to PredictSK—1 month after App Tracking Transparency went into effect and 1 week after SKAdNetwork was implemented. AppsFlyer is the first mobile attribution provider to deliver predictive analytics abilities.
Accelerating Development and Deployment on AWS
To build its novel solution, AppsFlyer relied on Amazon SageMaker and various AWS serverless services that shortened the path from research to production, such as AWS Lambda, a serverless compute service that lets companies run code without provisioning or managing servers. “The serverless architecture on AWS reduces development time and maintenance,” says Benjamin Winestein, senior software developer at AppsFlyer. “We can start with a small scale but then increase the scale with the confidence that everything will still work.”
PredictSK gives mobile app users a prediction score on a scale of one to nine. The solution uses Amazon SageMaker automatic model tuning—which finds the best version of a model by running many training jobs on the dataset—to automatically adjust thousands of combinations of algorithm parameters to rapidly improve the accuracy of predictive ML models. Each model then deploys; or if it needs additional training, AppsFlyer repeats the training flow using AWS Step Functions—a serverless function orchestrator that makes it simple to sequence AWS Lambda functions and multiple AWS services into business-critical applications. To maintain accuracy, each model is retrained monthly using tens of gigabytes of data.
AppsFlyer creates a custom ML model for every app that enters PredictSK. It does this for both security and user privacy reasons; no information is shared between apps. Additionally, AppsFlyer must train different ML models for different use cases because user behavior varies depending on the type of app. For example, users won’t behave on a healthcare app as they would on a gaming app. To meet that demand, AppsFlyer heavily relies on Amazon SageMaker multimodel endpoints, which provide a scalable, cost-effective way to deploy large numbers of custom ML models. AppsFlyer runs Amazon SageMaker on Amazon Elastic Compute Cloud (Amazon EC2) P3 Instances—which deliver high performance compute in the cloud, powered by NVIDIA V100 Tensor Core GPUs—and uses p3.2xlarge instances.
In the scheduled prediction flow, AppsFlyer uses Amazon SageMaker batch transform to run inferences every 1–24 hours on large batch datasets using a simple API. The prediction results are then stored in a database and made available to customers. “For a normal advertiser that uses no predictive or advanced insights, it would take around 30 days to receive any kind of LTV insights for a user,” says Hayet. “PredictSK cuts this time frame to as little as several hours.”
PredictSK’s near-real-time prediction flow operates on a serverless architecture and provides AppsFlyer’s customers with a predicted user benefit score nearly instantly—in 10–30 ms per inference from request to return. It processes several hundred gigabytes of user data each day, at tens of thousands of events per second, and will scale to hundreds of thousands of events per second. The solution consumes the day’s relevant events and saves that data to Amazon DynamoDB, a key-value and document database. It then uses AWS Lambda to prepare the data for inference and send it to an Amazon SageMaker multimodel endpoint. The inference results then are written to another table on Amazon DynamoDB to be sent to customers.
Optimizing a Novel Solution on AWS
AppsFlyer next plans to streamline—and potentially fully automate—the process of how user acquisition managers use PredictSK insights to optimize their campaigns. It is also exploring distributed training using Amazon SageMaker, which could help shorten training time and scale to support larger datasets.
On AWS, AppsFlyer quickly reacted to a significant industry change with a solution that improves campaign performance accuracy and protects consumer privacy while providing actionable insights earlier than other existing solutions. “We are happy to provide a kind of service that none of our competitors are offering our customers now,” says Hayet. “It’s an amazing position to be in, but we’re not resting. We keep progressing, improving, and working on this solution.” Levi adds, “Predictive analytics changes the mobile marketing game in a number of important ways, and we’re looking forward to offering the ecosystem more innovations moving forward.”
Founded in 2011, AppsFlyer provides software as a service for mobile-marketing analytics and attribution. Operating out of 20 global offices, AppsFlyer helps over 12,000 customers measure how users interact with brands through various apps, channels, and devices.
Benefits of AWS
- Implemented idea for ML algorithms in 6–8 months
- Produces insights based on the first 24–48 hours of user interaction
- Scales to hundreds of thousands of events per second
- Processes several hundred gigabytes of user data per day
- Retrains models monthly using tens of gigabytes of data
- Sees 10–30 ms per inference from request to return
- Predicts LTV in 1–24 hours compared to at least 30 days
- Cut time to production by 66% with the same number of staff
AWS Services Used
Amazon SageMaker helps data scientists and developers to prepare, build, train, and deploy high-quality machine learning (ML) models quickly by bringing together a broad set of capabilities purpose-built for ML.
AWS Lambda is a serverless compute service that lets you run code without provisioning or managing servers, creating workload-aware cluster scaling logic, maintaining event integrations, or managing runtimes.
AWS Step Functions
AWS Step Functions is a low-code visual workflow service used to orchestrate AWS services, automate business processes, and build serverless applications.
Amazon DynamoDB is a key-value and document database that delivers single-digit millisecond performance at any scale.
Companies of all sizes across all industries are transforming their businesses every day using AWS. Contact our experts and start your own AWS Cloud journey today.