Section: Overall Objectives
AVIZ's research on Visual Analytics is organized around four main Research Themes:
- Methods to visualize and smoothly navigate through large data sets
Large data sets challenge current visualization and analysis methods. Understanding the structure of a graph with one million vertices is not just a matter of displaying the vertices on a screen and connecting them with lines. Current screens only have around two million pixels. Understanding a large graph requires both data reduction to visualize the whole and navigation techniques coupled with suitable representations to see the details. These representations, aggregation functions, navigation and interaction techniques must be chosen as a coordinated whole to be effective and fit the user's mental map.
AVIZ designs new visualization representations and interactions to efficiently navigate and manipulate them.
- Efficient analysis methods to reduce huge data sets to visualizable size
Designing analysis components with interaction in mind has strong implications for both the algorithms and the processes they use. Some data reduction algorithms are suited to the principle of sampling, then extrapolating, assessing the quality and incrementally enhancing the computation: for example, all the linear reductions such as PCA, Factorial Analysis, and SVM, as well as general MDS and Self Organizing Maps. We investigate the generality of the approach and also explore other methods for cases such as for language processing where sampling severely reduces the quality of the result.
- Evaluation methods to assess their effectiveness and usability
Designing analysis components with interaction in mind has strong implications for both the algorithms and the processes they use. Some data reduction algorithms are suited to the following process: sampling, then extrapolating, assessing the quality and incrementally enhancing the computation. For example, all the linear reductions methods such as PCA, Factorial Analysis, and SVM, as well as general MDS and Self Organizing Maps can use that process. We investigate the generality of the approach and also explore other methods for cases such as for language processing where sampling severely reduces the quality of the result.
- Engineering tools
for building visual analytic systems that can access, search, visualize and analyze large data sets with smooth, interactive response.
AVIZ seeks at merging three fields: databases, data analysis and visualization. Part of this merging consists in using common abstractions and interoperable components. This is a long-term challenge, but it is a necessity because generic, loosely-coupled combinations will not achieve interactive performance.
Currently, databases, data analysis and visualization all use the concept of data tables made of tuples and linked by relations. However, databases are storage-oriented and do not describe the data types precisely. Analytical systems describe the data types precisely, but their data storage and computation model are not suited to interactive visualization. Visualization systems use in-memory data tables tailored for fast display and filtering, but their interactions with external analysis programs and databases are often slow.
These themes are presented separately, but they are closely linked: a good multi-scale visualization technique relies on an analysis method to generate the suitable data structure. The effectiveness of the Visual Analytics tool has to be evaluated at several levels (component, system, environment). Finally, to build Visual Analytics systems that manage large data sets, the software infrastructure has to provide the right abstractions and mechanisms. Therefore, each of the four research themes work together. One of the scientific challenges is to fit them all together into a coherent framework supporting the analyst's work process.