DF22400
Machine Learning (3p)
Introduction and Application for Automated Performance Tuning

EXERCISES FOR SELF-ASSESSMENT - HT/2012

Supervised Learning

• The complexity of most learning algorithms depends on the training set (size). Suggest a filtering algorithm that eliminates redundant instances.
• The sum-of-difference-squares used as error function is popular but not robust against outliers, and other error functions are possible. Suggest an error function that is robust against outliers.
• Show that the VC dimension of a line is 3.

Decision Trees

• In a univariate decision tree, numeric properties are often tested by a binary split (comparison of a single variable to a threshold value). Discuss the advantages and disadvantages of using ternary, quaternary etc. splits instead.
• Discuss advantages and disadvantages of using nonlinear split functions in multivariate decision trees, compared to linear ones (as presented).
• Extend the greedy decision tree creation algorithm of the lecture by backtracking to compute optimal decision trees for the given training set.