Ppdl improves user retention forecasting for paid channels
PPDL: A Real-world Industrial User Retention Ratio Forecasting Framework Integrating Physical Priors with Deep Learning
Machine Learning
Summary
Predicting how many users will keep using an app after joining through different ad channels is important for spending advertising money wisely. User retention tends to drop quickly at first and then stabilize, with seasonal ups and downs. The authors developed PPDL, a method that combines known mathematical patterns with deep learning to better predict these retention curves even when data is limited and varied across channels. Their experiments showed that PPDL is more accurate than current methods and works well on large real-world datasets.
What this means in practice
- •For digital marketing teams: Optimize advertising budgets by accurately predicting user retention ratios across different paid acquisition channels.
- •For platform growth analysts: Forecast user retention trends on short video platforms to guide strategic decisions on content and user engagement.
Authors
Zibo Zhao, Zhengxiong Guan, Chaoli Zhang, Linyuan Geng, Xuanbing Zhu, Zhonglong Zheng, Fan Wu
Abstract
In multi-channel paid user acquisition, early and accurate prediction of user retention at the channel level is crucial for optimizing budget allocation. User retention curves display a pronounced temporal pattern: an initial period of high churn transitions into long-term stability. This pattern is further characterized by regular fluctuations attributable to seasonality and exhibits high serial autocorrelation. These intrinsic properties make such curves highly suitable for analysis within a time-series forecasting framework. However, forecasting user retention ratio for large-scale short-video platform faces three major challenges: significant heterogeneity across channels, pronounced global trend of decay followed by saturation, and short look-back windows. To address these challenges, we propose PPDL, a novel forecasting framework that integrates physical priors with deep learning. We first introduce a trend-residual decomposition component. The trend is modeled using the Weibull distribution, whose parameters are learned via a Multilayer Perceptron (MLP). Secondly, for the residual component, we design an auxiliary embedding module on top of a deep learning backbone to maintain the channel identity awareness. Finally, to enhance the model's sensitivity to trends, we design a Multiscale Trend-penalized loss function. The proposed approach PPDL is validated through comprehensive experiments on industrial-scale datasets, covering three applications with an average of 30+ channels each. Experimental results show that PPDL achieves improvements across different backbones and significantly outperforms existing online solutions.