Applying Survival Analysis to User Activity Modeling: Concepts, Methods, and the KwaiSurvival Deep‑Learning Framework

With the rapid development of internet services, conventional north‑star metrics such as DAU no longer capture the full picture of user behavior because they ignore the timing of events. The article shows how survival analysis can incorporate event timing to provide a richer understanding of user activity.

Why use survival analysis for user activity? Traditional regression models focus only on binary outcomes, while survival analysis models both the occurrence and the time to event, allowing analysts to quantify how quickly users return to an app.

Survival analysis basics include definitions of event, survival time, censoring, and risk set, as well as typical data characteristics (non‑negative, discrete, no missing values). Important functions such as the survival function, hazard function, and cumulative hazard are introduced, with the Kaplan‑Meier (KM) estimator highlighted as a non‑parametric method.

Applying the concepts the article maps user‑level events (app launches) to survival analysis terminology, shows how to compute risk sets, and illustrates the construction of retention, risk, and survival curves. Visual examples compare regions with different activity patterns and demonstrate geographic clustering of user survival probabilities.

Modeling user activity involves fitting Cox proportional hazards models (including Anderson‑Gill counting processes for recurrent events) and evaluating them with concordance rather than AUC. Feature groups such as demographics, consumption behavior, and recommendation strategy are used, and SHAP values are employed to rank feature contributions both globally and locally.

KwaiSurvival framework is a Python‑based deep‑learning survival analysis library developed by Kwai. It integrates three models: DeepSurv (neural‑network Cox), Neural Multitask Logistic Regression (N‑MTLR), and DeepHit (enhanced with ResNet and rank loss). Comparative experiments on 70 W records with 74 features show that DeepHit outperforms traditional Cox in concordance while maintaining similar training time.

The source code and models are publicly available on GitHub (https://github.com/kwaiDA/KwaiSurvival/tree/kwaiDA-liuziyue), inviting the community to contribute additional survival‑analysis tools.