Abstract

One key challenge in data mining and manifold learning applications is the ability to manipulate datasets and to extract useful information. This triggered the need for dimensionality reduction techniques to help visualize and analyze the data without excess computation time. Several methods have been proposed in recent years that are based on the batch mode, meaning that all the data points need to be present during the run of the algorithm. Thus, any newly arriving point has to be added to the dataset and the algorithm has to be rerun on the entire set to produce the output. This limitation gave rise to the incremental form of some existing algorithms enabling them to adapt to newly added points without loss of computational time. Motivated by this incremental extension of the algorithms, this thesis addresses the limitations of some dimensionality reduction techniques and proposes incremental solutions. The core idea behind our proposed techniques is to enable these methods to adapt to the stream of data being added while preserving the significant characteristics of the low dimensional representation of the dataset. Our experimental results demonstrate an improved performance of the proposed approaches in comparison with existing algorithms