Immersive Interactive Data Mining and Machine Learning Algorithms for Big Data Visualization

The amount of collected Earth Observation (EO) images is increasing exponentially and their growth is currently in the order of several terabytes per day. Therefore, the ability to automatically store and retrieve these images based on their content is highly desired. Traditional approaches are not accurate and robust enough to handle this massive amount of data. However, the combination of artificial intelligence and human intelligence could deliver promising results. Therefore, this thesis addresses several challenges in the field of human-machine communication for data mining applications. This is mainly done by first introducing an Immersive Visual Data Mining (IVDM) system, including image collections and feature space visualizations, interactive dimensionality reduction, and active learning for image classification. A Cave Automatic Virtual Environment (CAVE) is employed to support the user-image interactions and also immersive data visualization, which allows the user to navigate through the images and explore them. The feature space is visualized by applying state-of-the-art dimensionality reduction techniques to reduce the dimensionality to 3D. Additionally, a novel algorithm based on Non-negative Matrix Factorization (NMF) is developed to arrange the images in 3D space by decreasing the occlusion among images and to make use of the display space more efficiently. Two interactive dimensionality reduction algorithms are introduced to enhance the discriminative property of the features by incorporating the user-image interactions. To annotate images, a novel active learning algorithm is proposed to choose the most informative images for labeling. Finally, experimental evaluations using publicly available data sets demonstrate the efficiency of the proposed algorithms.