9. Contributions

Throughout this thesis we have tried to note our contributions. This chapter is a short summary of all the contributions we mentioned, or should have mentioned. Our major contributions include

• Novel LR fitting procedure which uses CG to find an approximate solution the IRLS weighted linear least squares subproblem. This improves speed and handles data with linear dependencies.

• Demonstration of LR's suitability for data mining and classification applications on high-dimensional datasets.

• Publically available high-performance implementation of our IRLS algorithm, as well as the traditional CG-MLE. This software includes the CG implementation we wrote to support our LR software. Our CG implementation appears to be faster than the GNU Scientific Library's implementation.

• In-depth survey of IRLS and CG-MLE modifications, along with comprehensive empirical results.

In the course of our work, we also introduced a modification to the Cholesky decomposition to overcome linear dependency problems, described in Komarek and Moore [20]. However, LR methods using this decomposition were inferior to our final approach.