Survival Analysis Meets Data Stream Mining

Survival analysis deals with monitoring entities over their lifetime. The definition of "birth" and "death" events depends on the nature of a given entity. When we observe an infinite stream of birth and death events, at each point in time some of the monitored entities are "right-censored", i.e. we know the time elapsed since their birth event, but their death event has not occurred yet and we do not know when it will occur in the future. Often, the snapshots of partially censored observations keep arriving over time in the form of a data stream. Given each snapshot, we may be interested to predict the timing of death events for all live entities or, alternatively, to predict their label ("survived" or "failed") as a function of time. In this research, our intention is to modify standard classification algorithms, such as decision trees, so that they can seamlessly handle a snapshot stream of both censored and non-censored data. The objective is to provide reasonably accurate predictions after observing relatively few snapshots of the data stream and to improve the classification model with additional information obtained from each new snapshot.