Problem Statement:
With the significant improvements of our capacities on automatic data capturing and storage, large-scale
multi-modal data collections are available for achieving deeper understanding and facilitating better
decision-making in science, engineering, commerce, medicine and homeland security. For examples,
large-scale multi-modal data collections have led to significant improvements in healthcare diagnosis
and delivery. Unfortunately, it is not a trivial task for deriving useful knowledge and gaining significant
insights from large-scale
multi-modal data collections: (1) Most existing techniques for statistical data analysis
may not be scalable to the sizes of data collections, e.g., their computational complexity may increase exponentially
with the sizes of data collections; (2) The data items are normally represented
by high-dimensional multi-modal features and their statistical properties are heterogeneous, but most
existing techniques focus on single-modal data representation and implicitly assume that the statistical properties of the data items are
homogeneous in the high-dimensional feature space; (3) The data analysts are not the data
stakeholders, but most existing techniques for statistical data analysis have not provided a good
environment to enable visual-based communication between the data analysts
and the data stakeholders, thus they cannot share their interesting observations and understandable assessment
effectively. Visual analytics, which can seamlessly integrate statistical data analysis and visualization to
enable visual-based communication between the data analysts
and the domain experts, is very attractive for addressing these problems.
In order to develop new visual analytics frameworks for achieving better decision-making, there is an urgent need
of new mathematical foundations for multi-modal data transformation, knowledge discovery,
and interactive visualization for knowledge and hypotheses assessment.
Proposed Research:
In this project, we will focus on develop a novel framework to address the following issues more effectively:
(a) Multi-Modal Data Transformation and Dimension Reduction:
The data sets, which are collected from different sources may consist of multiple data types and can further be
represented by high-dimensional multi-modal
features. Thus it is very hard to manipulate the data items directly in the high-dimensional
multi-modal feature space and it is very important to transform the data items
into more suitable forms that can enable more effective data manipulation and exploration. Obviously,
such transformation should faithfully retain the rich contents of the original
data items (i.e., their statistical properties, their domain-dependent semantic interpretations,
their diverse similarity relationships, etc.).
(b) Visual-Based Knowledge and Hypotheses Assessment: In order to incorporate
visual analytics for improving data analysis and knowledge discovery, it is very important to
develop more effective visualization frameworks for assisting the data analysts and the domain experts on
sharing their interesting observations (i.e., useful knowledge and patterns) and assessing the
hypotheses for knowledge discovery.
Some pioneer works have been done by incorporating multivariate data analysis and multi-dimensional scaling
for supporting large-scale data visualization and exploration.
Even visualization can allow the data analysts and the domain experts to see large amount of data items at once,
visualizing large amount of data items on a size-limited display screen may seriously suffer from the
overlapping problem.
Because of the lack of suitable tools for adaptive data sampling and the shortage of a natural way to
support change of focus, all these existing techniques are unsuitable for
dealing with large-scale collections of data items. In addition, it is not a trivial task to obtain a good
similarity-preserving
projection of large amount of data items from the high-dimensional multi-modal feature space to a
two-dimensional
display space. To incorporate visualization for assisting the data analysts and the domain experts on
assessing the derived knowledge and the hypotheses for knowledge discovery, new frameworks should be
developed to achieve adaptive data sampling and similarity-preserving data projection.
If we know what we were doing, it wouldn't be research, would it?
---Albert Einstein(1879-1955)---