Katarzyna Woźnica
supervisor: Przemysław Biecek
The classical approach to train a machine learning model is to start with some default parameters and tune them for a selected dataset to maximize some measure of performance. And such process is repeated from scratch for each problem independently.
Meta-learning offers an alternative paradigm. This is a wide area of methods of systematically observing how different machine learning models perform on a wide range of learning tasks. Knowledge, which is extracted from some tasks, can be transferred to design optimal, predictive pipeline for new data.
This is an important problem as it may decrease training time, improve model generalisation and increase our knowledge about the training process.
It assumes that the expected performance of a model can be predicted based on various task's aspects called meta-features. The most frequently considered qualities are algorithm hyperparameters, statistical and information-theoretic dataset properties and model-based landmarkers trying to describe the relations between studied tasks in high dimensional space.
Existing approaches in meta-modeling are focused on searching for the best model but do not explain how these different aspects contribute to the performance of a model. To build a new generation of meta-models we need a deeper understanding of the relative importance of meta-features and construction of better meta-features.
In this talk, I will present novel technique to evaluate the relative importance of different groups of met.