Business Analytics Seminar: Mustafa Baydogan (UvA)
"Tree-based Representations for Time Series Data Mining"
Event details from: Business Analytics Seminar: Mustafa Baydogan (UvA)
- Date
- 9 January 2023
- Time
- 12:00 – 13:00
Abstract:
In the last decade, the increasing use of temporal data, especially time series data, has initiated a great deal of research and development attempts. Many data sources in different fields, such as in medicine, education and finance, naturally generate time series data (e.g. electrocardiogram (ECG), weekly sales totals, and prices of stocks) and modeling of this data has become important to understand the system generating the time series. Modeling time series data involves various tasks such as forecasting, discovery of underlying driving factors or variables, or understanding the dynamics of the series in terms of its variance and other characteristics. A fundamental problem in time series analysis approaches is how to represent/model the time series data. However, modeling the time series data is difficult when:
- There are hundreds (to thousands) or more variables with thousands (to millions) of observations.
- They are of different type (e.g. numerical, ordinal and nominal).
- For multivariate data, the observations are sampled in nonuniform time points and/or the sampling frequency for each variable may not be the same.
- The relations between the variables are nonlinear. Interaction effects are to be detected, and redundant variables are expected to be linked.
- There is an uncertainty component to be modeled. Outliers and missing values are common.
- The temporal and/or spatial effects, and transient effects are important.
This talk is about the use of tree-based algorithms for modeling of the time series under aforementioned critical challenges. Novel tree-based ensemble learning strategies are developed to handle a number of statistical learning tasks for time series directly. Developed strategies provide a research direction for approaches that break from the linear dependency models to potentially foster other promising nonlinear approaches. Moreover, proposed research based on the tree-based learning is flexible in handling any type of sequential data from different applications. Our experiments on sequential data from diverse domains have shown the success of the proposals. In time series data mining domain, there are few studies proposing a generalized view to alternative statistical learning tasks such as forecasting, similarity search, classification and clustering. The most important contribution of this research is the introduction of the generalized framework that can handle multiple important time series data mining tasks under a unified framework.
General information
This seminar will be organised in a hybrid setup. If you are interested in joining this seminar, please send an email to the secretariat of Amsterdam Business School at secbs-abs@uva.nl.
