Unlocking revenue: newcomer classification for boosted profits
summary
Built a multi‑task neural model with hard parameter sharing that takes ~300 features from a 1‑day observation window to jointly predict churn probability (BCE) and LTV (log‑normal loss), calibrated with Gini and Spearman and tuned to avoid underestimates. Deployed through an MLOps inference pipeline that selects high‑probability whales for exclusive free‑gift offers; AB tests report a >4% LTV uplift. Synthetic oversampling is not used.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 0:45Meet the speaker
- 3:19New user classification problem
- 5:05Lifetime value challenges
- 8:06Predictive analytics modeling
- 11:46Multitask learning architecture
- 13:42Loss function for LTV
- 15:52Neural network design
- 16:35Model performance metrics
- 17:57Model calibration
- 18:39Deployment pipeline
- 19:45Key takeaways
- 20:06Q&A
related talks
19mUnveiling the Potential: Exploring Upselling, pLTV, and ML for Operations
27mPreventing churn like a bandit
23mCustom-made games with machine learning and big data
32mClassifying the variety of customers online engagement for churn prediction
13mOptimizing Subscription Services with Uplift Modeling
01h 00mHow to win at monetizing games as a service
30mWinning (Some) Churners Back
32mWhy start using uplift models for more efficient marketing campaigns