Insights into Machine Learning: Data Clustering and Classification Algorithms for Astrophysical Experiments

Data analysis domain dealing with data exploration, clustering and classification is an important problem in many experiments of astrophysics, computer vision, bioinformatics etc. The field of machine learning is increasingly becoming popular for performing these tasks. In this thesis we deal with machine learning models based on unsupervised and supervised learning algorithms. In unsupervised learning category, we deal with Self-Organizing Map (SOM) with new kernel function. The data visualization/exploration and clustering capabilities of SOM are experimented with real world data set problems for finding groups in data (cluster discovery) and visualisation of these clusters. Next we discuss ensembling learning, a specialized technique within the supervised learning field. Ensemble learning algorithms such as AdaBoost and Bagging have been in active research and shown improvements in classification results for several benchmarking data sets. They grow multiple learner algorithms and combine them for getting better accuracy results. Generally decision trees learning algorithm is used as base classifiers to grow these ensembles. In this thesis we experiment with Random Forests (RF) and Back-Propagation Neural Networks (BPNN) as base classifiers for making ensembles. Random Forests is a recent development in tree based classifiers and quickly proven to be one of the most important algorithms in the machine learning literature. It has shown robust and improved results of classifications on standard data sets. We experiment the working of the ensembles of random forests on the standard data sets available in University of California Irvine (UCI) data base. We compare the original random forest algorithm with their ensemble counterparts and discuss the results. Finally we deal the problem of image data classification with both supervised (ensemble) learning and unsupervised learning. We apply the algorithms developed in the thesis for this task. These image data are taken from the MAGIC telescope experiment, which collects the images of particle rays coming from the outer universe. We apply the ensembles of RF, BPNN for making a supervised classification of images and compare the performance results. Then we discuss a SOM system, developed for making an automatic classification of images using the unsupervised techniques.