Predictive modeling using sparse logistic regression with applications

TL;DR: It is shown that a combination of a careful model assessment scheme and automatic feature selection by means of logistic regression model and coefficient regularization create a powerful, yet simple and practical, tool chain for applications of supervised learning and classification.

Abstract: In this thesis, sparse logistic regression models are applied in a set of real world machine learning applications. The studied cases include supervised image segmentation, cancer diagnosis, and MEG data classification. Image segmentation is applied both in component detection in inkjet printed electronics manufacturing and in cell detection from microscope images. The results indicate that a simple linear classification method such as logistic regression often outperforms more sophisticated methods. Further, it is shown that the interpretability of the linear model offers great advantage in many applications. Model validation and automatic feature selection by means of `1 regularized parameter estimation have a significant role in this thesis. It is shown that a combination of a careful model assessment scheme and automatic feature selection by means of logistic regression model and coefficient regularization create a powerful, yet simple and practical, tool chain for applications of supervised learning and classification.