Discovering Task Neighbourhoods Through Landmark Learning Performances

Arguably, model selection is one of the major obstacles, and a key once solved, to the widespread use of machine learning/data mining technology in business. Landmarking is a novel and promising metalearning approach to model selection. It uses accuracy estimates from simple and eÆcient learners to describe tasks and subsequently construct meta-classi ers that predict which one of a set of more elaborate learning algorithms is appropriate for a given problem. Experiments show that landmarking compares favourably with the traditional statistical approach to meta-learning.