Event Prediction: Learning from Ambiguous Examples
Gary M. Weiss, Haym Hirsh · 1998
Event prediction is an important problem with many realworld applications. For the majority of these applications, it is not necessary to predict the exact time an event will occur---it is acceptable to predict that the event will occur within some time interval. The use of this time interval introduces ambiguity into the event prediction problem, permitting it to be viewed as a multiple-instance learning problem. We have developed timeweaver, a genetic-algorithm based learning system that is capable of solving these event prediction problems. Timeweaver handles the ambiguity in the learning problem by employing a special evaluation function. In this paper we also describe how some of the ambiguity can be eliminated by reformulating the learning problem. Introduction There are many situations where one would like to learn to predict the occurrence of a specific type of event, a target event, before it actually occurs. Representative applications include predicting telecommunication ...