Learning from complex event interval sequences

PhD student: Zed Lee, DSV

Expert reviewers: Toon Calders, University of Antwerp and Sindri Magnússon, DSV

Main supervisor: Panagiotis Papapetrou, DSV

Supervisor: Tony Lindgren, DSV


As modern systems are getting more complex, the amount of temporal data is emerging at a rapid rate. Such data can be found in various data domains, such as sign language, healthcare diagnosis data, and manufacturing logs. Hence, it is important to develop novel methods and techniques for extracting patterns from this flooding amount of complex data. Event sequences can be a great way to analyze complex temporal data, by relaxing their complexity to a discretized form. Examples of benefits from doing so include reducing random noise in data, avoiding problems with sampling data at different frequencies and different time units, summarizing large numbers of data in a concise way, and helping with algorithms to cope with missing data. This thesis first introduces a set of methods for knowledge extraction from event interval sequences by detecting patterns of interest. The pattern of interest can be a set of frequent event intervals and their pairwise relations, a set of event intervals showing the most disproportional proportion, and a set of different time series variables showing similar trends over a specific time range. This thesis then applies temporal abstractions to a variety of temporal data types and application domains such as histogram snapshots of truck sensors, sales data with different product groups, and general multivariate time series to capture the temporal relations among different variables.