From insights to innovations: data mining, visualization, and user interfaces
Johan Himberg · Aaltodoc (Aalto University) · 2004
This thesis is about data mining (DM) and visualization methods for gaining insight into multidimensional data. Novel, exploratory data analysis tools and adaptive user interfaces are developed by tailoring and combining existing DM and visualization methods in order to advance in different applications. The thesis presents new visual data mining (VDM) methods that are also implemented in software toolboxes and applied to industrial and biomedical signals: First, we propose a method that has been applied to investigating industrial process data. The self-organizing map (SOM) is combined with scatterplots using the traditional color linking or interactive brushing. The original contribution is to apply color linked or brushed scatterplots and the SOM to visually survey local dependencies between a pair of attributes in different parts of the SOM. Clusters can be visualized on a SOM with different colors, and we also present how a color coding can be automatically obtained by using a proximity preserving projection of the SOM model vectors. Second, we present a new method for an (interactive) visualization of cluster structures in a SOM. By using a contraction model, the regular grid of a SOM visualization is smoothly changed toward a presentation that shows better the proximities in the data space. Third, we propose a novel VDM method for investigating the reliability of estimates resulting from a stochastic independent component analysis (ICA) algorithm. The method can be extended also to other problems of similar kind. As a benchmarking task, we rank independent components estimated on a biomedical data set recorded from the brain and gain a reasonable result. We also utilize DM and visualization for mobile-awareness and personalization. We explore how to infer information about the usage context from features that are derived from sensory signals. The signals originate from a mobile phone with on-board sensors for ambient physical conditions. In previous studies, the signals are transformed into descriptive (fuzzy or binary) context features. In this thesis, we present how the features can be transformed into higher-level patterns, contexts, by rather simple statistical methods: we propose and test using minimum-variance cost time series segmentation, ICA, and principal component analysis (PCA) for this purpose. Both time-series segmentation and PCA revealed meaningful contexts from the features in a visual data exploration. We also present a novel type of adaptive soft keyboard where the aim is to obtain an ergonomically better, more comfortable keyboard. The method starts from some conventional keypad layout, but it gradually shifts the keys into new positions according to the user's grasp and typing pattern. Related to the applications, we present two algorithms that can be used in a general context: First, we describe a binary mixing model for independent binary sources. The model resembles the ordinary ICA model, but the summation is replaced by the Boolean operator OR and the multiplication by AND. We propose a new, heuristic method for estimating the binary mixing matrix and analyze its performance experimentally. The method works for signals that are sparse enough. We also discuss differences on the results when using different objective functions in the FastICA estimation algorithm. Second, we propose "global iterative replacement" (GIR), a novel, greedy variant of a merge-split segmentation method. Its performance compares favorably to that of the traditional top-down binary split segmentation algorithm.