Machine Learning in Dynamic Data Environments

Course goals

The aims of this course:

• familiarize with current literature and developments in data stream mining / online machine learning.
• understand Nonstationary Environments: Develop a profound understanding of nonstationary data, including concept drift, covariate shift, and evolving data distributions.
• Adaptive Models: understand the principles and methods behind adaptive machine learning/data stream mining techniques, such as change detection, forgetting, windowing, and adaptive ensembles.
• analyze the relation between different categories of change, processing scenarios and adaptive techniques.
• evaluate the performance of adaptive techniques and reflect on limitations of exemplary approaches.

Assessment
This research-oriented course integrates formative self-assessment and graded assessments of both, understanding of theory and practical skills.
Throughout the course, there will be given opportunities for self-assessment. These are designed to provide immediate feedback that guides through the learning process and prepare for the later graded assessment.

The graded assessment is done by a written exam.

To qualify for a repair of the final result the mark needs to be at least a 4, or “AANV”.

Content

The world is dynamically changing and non-stationary. Among the various types of change that a machine learning system faces in such a dynamic environment, non-stationary data distributions are a common and important challenge.
This phenomenon is denoted as concept or population drift, and has led to the development of dedicated subfields in machine learning and statistics, such as data stream mining, online machine learning, change detection, or change mining/change understanding.

This research-oriented course is designed as advanced course on machine learning/data mining for students with a good knowledge of traditional (stationary, offline) machine learning/data mining techniques, in particular in classification and clustering.

Course content:

•   categorizing change: Non-stationarity, concept drift, population drift, shift, concept evolution
•   key principles and techniques, such as change detection, windowing, forgetting; chunk-based and one-by-one processing
•   adaptive/online learning methods, in particular for Data Stream Classification, such as Hoeffding Trees or Adaptive Ensembles
•   evaluation and monitoring methodology for non-stationary environments, like out-of-time-evaluation, prequential evaluation, recovery analysis
•   understanding change: Interpretability and explainability in non-stationary environments
•   exemplary frameworks for Data Stream Mining, like RIVER/scikit-multiflow and Massive Online Analysis (MOA)