Towards AutoML in the presence of Drift
Jorge Madrid, Hugo Jair Escalante, Eduardo Moales · 2018
In many real-world scenarios AutoML systems are necessary due to the limited expertise in Machine Learning by developers. Research in this field has lead to the development of state-of-the-art solutions for supervised Machine Learning tasks that work effectively in different application areas. However, these don’t consider the changing nature inherent to several real-world situations (e.g. recommendation systems, spam-filtering). We briefly describe a first attempt to cope with such situation. We extended Auto-Sklearn with intuitive mechanisms to allow it to cope with data distributions that change over time. Results demonstrate the effectiveness of the proposed methodology. An extended version of this work was presented in AutoML Workshop @ ICML2018.