•   When: Friday, December 08, 2017 from 02:00 PM to 04:00 PM
•   Speakers: Xing Wang
•   Location: ENGR 2901
•   Export to iCal

The problem of efficient discovery of previously unknown, frequently appearing patterns in a time-series dataset has received much attention in the past decade. It has been used in many data mining tasks such as rule discovery, novelty detection, clustering and summarization. A related task is the timely detection of anomalies in time series. Anomalies convey critical and actionable information in applications such as health care, equipment safety, security surveillance, and fraud detection. Consequently, the problem of anomaly detection has also been studied in diverse research areas.

Most existing work on patterns discovery in time series requires prior knowledge to be set by the user, or focuses on univariate data. For example, one approach to classify time series data is to find meaningful patterns in time series that can differentiate between classes. However, most techniques that find such patterns require the user to specify the length of the patterns. There are two limitations with this requirement in real world problems. First, the user may not know the exact pattern length, or even the best range of lengths in advance. Second, restricting the discovery to only fixed length patterns limits the algorithm’s exploratory capacity since multiple patterns of different lengths may co-exist in a time series.