A Fuzzy-Based Approach to Survival Data Mining

Abstract

Traditional data mining algorithms assume that all data on a given object becomes available simultaneously (e.g., by accessing the object record in a database). However, certain real-world applications, known as survival analysis, or event history analysis (EHA), deal with monitoring specific objects, such as medical patients, in the course of their lifetime. The data streams produced by such applications contain various events related to the monitored objects. When we observe an infinite stream of events, at each point in time (the “cut-off point”), some of the monitored entities are “right-censored”, since they have not experienced the event of interest yet and we do not know when the event will occur in the future. In snapshot monitoring, the data stream is observed as a sequence of periodic snapshots. Given each snapshot, we are interested to estimate the probability of a critical event (e.g., patient death or equipment failure) as a function of time for every monitored object. In this research, we use fuzzy class label adjustment so that standard classification algorithms 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 performance with additional information obtained from each incoming snapshot. The proposed fuzzy-based methodology is evaluated on real-world snapshot streams from two different domains of survival analysis.