Begin: Thursday, 18. of July 2019 (6:15 PM)
Location: Warburger Str. 100, Lecture Room O2
Abstract:
Many real world problems involve a temporal or probabilistic predictive element, e.g., predicting a cancer patient's time-to-event distribution, forecasting tomorrow's weather, or predicting failure risk of manufacturing equipment. Machine learning toolboxes such as sklearn, mlr, Weka provide unified interfaces to a range of model classes and powerful model composition strategies such as automated tuning and pipelining - though the state-of-art mostly covers point prediction tasks on single data frame or single matrix container, while support for probabilistic prediction, prediction intervals, or temporal learning tasks is quite limited.
In this talk, we give an introduction to key features and software architecture principles of ML toolboxes, with a specific view towards semantic and architectural expressivity necessary for probabilistic and temporal learning, supplemented by case study experience from the skpro (probabilistic prediction), sktime (learning with time series) and MLJ (machine learning in Julia) toolbox projects.
This will be joint with a gentle introduction to supervised learning methods which produce prediction ranges, and a taxonomy of time series related learning tasks, from a general black-box viewpoint - accompanied by an overview of higher-order model building strategies such as composition and reduction, and best practice advice on how to evaluate and compare performance of models fairly, or Bayesian and frequentist methods on equal footing.
Depending on audience interest, demos of recent and current projects may be presented at the end:
https://github.com/alan-turing-institute/sktime
https://github.com/alan-turing-institute/skpro
https://github.com/alan-turing-institute/distr6
https://github.com/alan-turing-institute/mlj.jl
