Getting Ready for Change: Handling Concept Drift in Predictive Analytics

Опубликовано 27 июля 2016, 2:43

In the real world data often arrives in streams and evolves over time. Concept drift in supervised learning means that the relation between the input data and the target variable changes. Therefore, in many real-world applications the learning models need to adapt to the anticipated changes. In this talk I will overview the state of the art in concept drift research in data mining and related areas. First, I will introduce the problem of concept drift with illustrative real-world examples, characterize adaptive learning process, categorize existing strategies for (reactive) handling concept drift in the most assumed setting � unpredictable changes happen in hidden contexts that are not observable to the adaptive learning system. Then, I will show why from the application perspective it is interesting to look into several other operational settings that commonly occur in practice, but have been underexplored in academia. In particular, I will show that there is a room for proactive approaches for handling. I will conclude the talk with an overview of the recent trends and next challenges in concept drift research.