Browsing by Author "Thom, Dennis"

Now showing 1 - 2 of 2

Open Access
Dynamic ontology supported user interface for personalized decision support
(2012) Bosch, Harald; Thom, Dennis; Heinze, Geoffrey-Alexeij; Wokusch, Stefan; Ertl, Thomas
European citizens are increasingly aware of the influence of air quality and weather on their health and quality of life. At the same time, more environmental information is freely available through a plethora of websites, dedicated portals, and web services. In order to exploit these data for personal decisions one has to identify, retrieve, and combine the information that is relevant to one's personal situation, planned activity, and information need. Often, this task is hindered by different data formats, display styles and data resolutions. The PESCaDO system is a web-based decision support system addressing this issue. The inquiry to the system, as well as the system's result, can cover a broad range of environmental aspects and personal situations and is therefore quite complex. In this work we present a novel approach on how the system can actively assist users in all steps of the decision making process, especially by enhancing the user interaction. This approach combines an intelligent dialog steering method based on analyzing the domain ontology with flexible, dynamic data visualizations for a situation depending orchestration of data sources. Both aspects have been evaluated in on-line user studies, as well as with an expert evaluation of the whole system.
Open Access
Visual analytics of social media for situation awareness
(2015) Thom, Dennis; Ertl, Thomas (Prof. Dr.)
With the emergence of social media services and other user-centered web platforms the nature of the modern internet changed substantially. While it has since been a vast source of information and news on all kinds of topics, it recently grew into a continuous stream of knowledge, observations, thoughts, and situation reports. They are provided in real-time by millions of people from all over the world. This change also offers completely new possibilities for domains that rely on good situation awareness, such as disaster management, emergency response, disease control, and several forms of command and control environments. Analysts can find eyewitness videos of ongoing critical events in Youtube, they can observe the movement and communication behavior of Facebook users during evacuation measures, and they are enabled to trace the outspread of an epidemic disease just by highlighting symptom related keyword usage in Twitter. However, the data sizes that need to be processed in order to identify relevant entries, produce comprehensible overviews, and detect anomalous patterns pose one of the most challenging analytics problems of our time. Not only the volume of data generated on a daily basis is larger than any other single database from the pre-internet era. The data is furthermore streamed in real-time at substantial velocity; it comes in a great variety, including text snippets, images, videos and network information; and it contains inaccuracies, misleading information, rumors, and fake meta-data, leading to uncertain veracity. In contrast to most other computer science challenges, social media analytics thus fully covers all characteristics that have been commonly referred to as the "four V's" of big data. By tightly integrating approaches from the areas of data mining, information retrieval, natural language processing, human computer interaction, and data visualization the emerging field of visual analytics has been devised to tackle these challenges. As a descendant of the more general field of information visualization, visual analytics strives to merge the strengths of highly interactive visual interfaces with the computational power of automatic statistical algorithms. The goal of this combination is to advance problem solving in areas where a human analyst alone would be overwhelmed by the data volumes, while, at the same time, sheer processing power alone would not enable analysts to identify underlying patterns and relate information to semantic knowledge. This thesis identifies four visual analytics requirements that have to be addressed to allow comprehensive situation awareness based on social media: Access to data, visualization of context, coping with semantic complexity, and scalable processing. Based on core ideas of visual analytics, this work contributes three distinct techniques that allow to tackle access, context, and complexity, as well as a prototypical implementation that integrates all of them and allows scalable processing of the data. Means of iterative query optimization and hierarchical exploration of data samples are presented that allow to cope with the problem of rate limited web data collection. The challenge of relating information to space, time, and context is solved by a novel technique that automatically detects and visually highlights possibly relevant events. Here, a sophisticated language model based on large volumes of data is employed to separate meaningful and related information from signal noise. Finally, the possibility to drill-down into complex topics and to enable ongoing situation monitoring is achieved by means of interactive classifier training and orchestration. The thesis furthermore presents an overarching analytics model, which integrates all solutions and relates their distinct capabilities. %All of the presented methods combine the analytical power of data mining and information retrieval algorithms with the capabilities of human cognition by means of highly interactive visual interfaces. The techniques, their prototypical implementation, as well as the overarching analytics model are thoroughly evaluated, and they are compared with other approaches in context of the relatively young scientific discourse. Along these lines, it is demonstrated how the aspects of user-driven detection and data-driven discovery distinctly align with supervised and unsupervised methods in machine learning. From the lessons learned, it is conclusively shown that visual configuration and steering of supervised classification on one hand, and the enhancement of visual interfaces through unsupervised clustering on the other hand, are two complementary concepts embedded at the very heart of visual analytics. The presented overarching analytics model might help to further enhance previous definition approaches and ostensive conceptions existing in the field.