Developing library of tree-based models for survival analysis

Abstract

One of the main tools for describing and predicting the occurrence of events is survival analysis, which allows you to predict not only the probability and time of events but also the changing of probability over time. This article presents Survivors, an open-source Python library that helps solve problems of survival analysis, build individual forecasts of survival and risk functions, investigate data dependencies, evaluate the quality of forecasts, and conduct experimental studies. The library uses new methods of constructing tree-based models of survival analysis with high sensitivity to real datasets. In particular, the paper presents a new histogram approach for searching the best split of censored data. The models can handle categorical and missing values, cases of informative censorship, and multimodal time distribution. The paper describes the architecture and components of the library, the features of the software implementation, and an experimental comparison with existing libraries of survival analysis.