Continuously Predicting the Completion of a Time Intervals Related Pattern.

In various domains, such as meteorology or patient data, events’ durations are stored in a database, resulting in symbolic time interval (STI) data. Additionally, using temporal abstraction techniques, time point series can be transformed into STI data. Mining STI data for frequent time intervals-related patterns (TIRPs) was studied in recent decades. However, for the first time, we explore here how to continuously predict a TIRP’s completion, which can be potentially applied with patterns that end with an event of interest, such as a medical complication, for its prediction. The main challenge in performing such a completion prediction occurs when the time intervals are coinciding, but not finished yet, which introduces an uncertainty in the evolving temporal relations, and thus on the TIRP’s evolution process. In this study, we introduce a new structure to overcome this challenge and several continuous prediction models (CPMs). In the segmented CPM (SCPM), the completion probability depends only on the pattern’s STIs’ starting and ending points, while a machine learning-based CPM (CPML) incorporates the duration between the pattern’s STIs’ beginning and end times. Our experiment shows that overall, CPML based on an ANN performed better than the other CPMs, but CPML based on NB or RF provided the earliest predictions.