Continuously predict user retention in the crucial first seconds and minutes after a new user onboards.
Survival analysis is one of the most developed fields of statistical modeling, with many real-world applications. In the realm of mobile apps and games, retention is one of the initial focuses of the publisher once the app or game has been launched. And it remains a significant focus throughout most of the lifecycle of any endeavor.
The inverse of retention is the churn rate. Stated simply, how many new users remain on board after a given period? Usually, this is measured in cohort data analysis across days, weeks or months. For some segments of the mobile gaming industry, the average for 1st day churn hovers at a mind-boggling 70%. This is not a new benchmark; there are endless SaaS vendors, blogs and other resources and solutions that address this problem.
However, they seem to focus on dealing with 1st day churn only after it has occurred, where the price to re-engage and reactivate a churned user is already increasing. This is in addition to the even greater costs of acquiring new users daily to compensate for the persistent loss.
Some companies go the extra step of combating churn in real time. They will often conduct a delicate exercise using programmatic rules set by product managers and analysts to segment users into groups based on similar attributes. They will then engage with the users according to their behavior patterns. Machine learning models are then added to the process, resulting in pipelines that include both real-time segmentation and user predictions, with optimized retention-based actions ready on call.
After speaking with many industry professionals, I understand that it’s not the lack of academic research or data science prowess that’s impeding our advancement in churn-rate improvement. What’s crucial is a rapid deployment process with the right ‘infrastructure’, called MLOps. Such a solution will only require minimal effort from in-house data engineers and can drive the business from lab to production in weeks rather than many months.
A data scientist working for an app/gaming company can now produce a real-time churn prediction pipeline consisting of multiple streaming data sources beyond the in-app user behavioral events. They can include marketing attribution data and even external data streams such as weather, local news and social media. Furthermore, given our ‘new normal’, monitoring of the freshly minted microservice for data drift can automate the triggering of retraining and deployment workflows. Testing, validation and even Canary rollouts and A/B testing are also included OOB (out of the box).
The challenge of addressing the user’s probability to be retained can now be realized. Taking only weeks to get ML into production. We can finally move up the ladder of proof to prescriptive analytical solutions that account for not just ML-powered micro segmentations and user retention but also for the optimization of the proposed action (type, size, creative, medium) with best LTV and ROI. Iguazio's comprehensive data science platform enables data scientists to focus exactly on this - the business logic — without having to deal with the infrastructure overhead. It was built with enterprise-scale and real-time requirements from the inception. Using open-source frameworks creating a fully automated, low latency, auto-scaling, inferencing endpoint is no longer a developer, data engineering, DevOps, IT, PM roadmap or resource problem. The data scientist can focus on solving the business problem, while the MLOps headache has been abstracted.
Watch this webinar with Product Madness and Iguazio to learn more about Predicting 1st-Day Churn in Real-Time: