Mining frequent and top-K High Utility Time Interval-based Events with Duration patterns

Traditional frequent sequential pattern mining only considers the time point-based item or event in the patterns. However, in many application, the events may span over multiple time points and the relations among events are also important. Time interval-based pattern mining is proposed to mine the interesting patterns of events that span over some time periods and also by considering the relations among events. Previous works of time interval-based pattern mining focused on the relations between events without considering the duration of each event. However, the same event with different time duration may cause different results. In this work, we propose two algorithms, SARA and SARS, for mining frequent time interval-based events with duration, TIED, patterns. TIED patterns not only keep the relations between two events but also reveal the time periods when each event happens and ends. For the performance evaluation, we propose a naive algorithm and modify a previous algorithm along with the implementation of SARA and SARS. The experimental results show that SARA and SARS are more efficient in terms of execution time and memory usage than other two algorithms. Moreover, we extend this work by considering utility value or importance of event in each time stamp. Therefore, we propose another new High Utility Time Interval-based Events with Duration, HU-TIED, pattern. HU-TIED incorporates the concept of utility pattern mining and TIED pattern mining. We design an algorithm, LMSpan, to mine top-K HU-TIED patterns. For the performance evaluation, we design a baseline algorithm, GenerateNCheck to compare with LMSpan. LMSpan takes less time and memory and generates less candidates than GenerateNCheck.