Towards Efficiently Mining Frequent Interval-Based Sequential Patterns in Time Series Databases

Nowadays time series mining has been taken into account in many various application domains. One of the most popular mining tasks in the existing works is the frequent pattern mining task on time series databases. Periodic patterns in a time series are often examined in this task. Such patterns help us understand more the corresponding object observed on a regular basis. As we extend our consideration to a group of many different objects to find out their common behaviors repeating over time, we need a pattern type to be more informative and thus a solution to discover the hidden patterns. Therefore, our work aims at so-called interval-based sequential patterns frequently in a time series database. We also provide two different solutions to mining such frequent patterns: the first one based on the existing ARMADA solution with the additional preprocessing and post-processing and the second one based on our new FITSPATS algorithm with the use of stems as suffix expansion and a temporal pattern tree. Experimental results have shown that our solutions are capable of discovering the frequent interval-based sequential patterns in a time series database and the FITSPATS algorithm is more effective and efficient for the task.