'Data Visualization and Data Analysis'

Task Force of IEEE CIS Technical Committee on Data Mining

Organizers:

Barbara Hammer, Bielefeld University

Laurens van der Maaten, Delft University of Technology

Daniel Keim, University of Konstanz

!!! Current events !!!

• Workshop Visual Analytics using Multidimensional Projections at Eurovis 2013
• Workshop NC² in Saarbrücken with a special focus topic on learning interpretable models
• Workshop VR/AR der GI
• Summer School on Interactive Intelligent Spaces
• Special Issue on Machine Learning for Remote Sensing Data Processing
• Workshop on High Dimensional Data Mining at IEEE ICDM

Past events

Background:

The rapidly increasing availability of electronic data in virtually all aspects of daily life confronts people with huge amounts of potentially valuable information which is often not easily accessible due to its sheer size and complexity. In this context, data visualization plays a key role to establish an intuitive interface to the data such that humans can directly rely on their astonishing cognitive capabilities for visual perception. Data visualization constitutes a matured field of research with widespread applications in diverse areas such as medical data analysis, text mining, bioinformatics, geoinformatics, scientific visualization, etc., and dedicated topics of data visualization have been discussed in numerous contributions to international conferences and journals. This is accompanied by open source and commercial software packages which allow users to directly visualize data. In the context of machine learning and data mining numerous algorithms have recently been developed: popular dimensionality reduction techniques of (mostly high dimensional vectorial) data which replace a given set of data points by low-dimensional vectors and display these on the screen include, for example, t-SNE, Isomap, Isotop, SOM, XOM, MDS, LLE, ...; recent efficient clustering techniques such as affinity propagation, spectral clustering, or relational neural gas allow to deal with general dissimilarity data; further developments which are important for the proposed topic are topic modeling such as latent Dirichlet allocation, developments in collaborative filtering such as weighted SVD, or methods which allow an efficient inspection of receptive fields or similar such as LVQ, ICA, etc. Typically, classical information visualization as considered e.g. in connection to computer graphics goes far beyond data projection and deals with dedicated techniques to display information in appropriate images or video streams. Interestingly, while many dimensionality reduction techniques have been developed in the field of data mining and machine learning, more advanced information visualization tools are widely unknown to the CIS community. Conversely, many recent dimensionality reduction and data mining techniques are not yet sufficiently known in classical information visualization. One goal of the proposed task force is to bridge this gap between the communities.

Further, the rapid technological development in terms of sensor technology, dedicated data formats, and automated data storage, pose new challenges for data visualization: methods have to deal with very large data sets such that only tools which rely on finite memory and at most linear time are still affordable; rather than simple euclidean vectors, data displays very high dimensionality or even more complex structures, such that methods have to cope with specific (probably non-metric) dissimilarity measures or dedicated data formats; often, a direct projection of data to low dimensions is no longer sufficient for adequate data visualization, rather dedicated data mining tools have to be integrated such as prior clustering, feature selection, or information extraction; users of data visualization tools are usually not experts in the field of data visualization (or even computer science), such that there is a need for parameterless methods and an automatic selection of the best method for the given problem - currently, users often still rely on in general unsatisfactory methods such as euclidean MDS simply because of its easy availability; data visualization per se constitutes an ill-posed problem and it is easily possible to visualize aspects of the data which are not relevant for the specific application or which are even due to pure noise within the data, hence it is necessary to choose an appropriate bias of what to visualize; since the objective of data visualization is not clear a priori nor universal for any given task, the question of how methods can be evaluated is an important unsolved problem. These issues pose quite a few problems to data visualization which will be addressed in the proposed task force.

Mission and Topics:

• disseminate the existing matured data visualization and analysis tools, and make end-users and researchers aware of the state-of-the-art in data visualization and analysis,
• bridge the gap of dimensionality reduction techniques as developed in the context of machine learning and advanced information visualization as represented e.g. by the IEEE CS Technical Committee on Visualization and Graphics,
• to identify new research challenges in the field which arise from modern data formats and data sets for visualization, thereby also bridging the gap between dimensionality reduction techniques as developed in the field of machine learning and advanced data mining techniques,
• to help in establishing standards for evaluation and benchmarking.

• How can methods be adapted to deal with very large data sets?
• How can specific data formats be visualized properly and efficiently such as spatio-temporal data, spectral data, data characterized by a general dissimilarity measure, etc.?
• How can we visualize similarity data which are inherently non-metric such as semantic similiarities? How can we visualize data that exhibits different types of objects (and possibly also missing values), such as user-movie ratings? How can we construct visualization techniques that do not operate on vectorial data or Euclidean distances, but on types of data that data analysts often encounter nowadays?
• What are interactions of data visualization and standard data mining technology such as clustering, information extraction, etc.? What is the connection of dimensionality reduction techniques for direct data display and dedicated methods of information visualization?
• What are reasonable biases of data visualization to shape this inherently ill-posed problem? Can this be controlled by the user in an efficient way?
• How can data visualization be evaluated? What are reasonable benchmarks? What are reasonable evaluation measures?
• One way to inspect why two points are similar with respect to some visualization is by back-projecting the visualization into the data space. Can we devise new techniques to do such back-projections? How can we evaluate the quality of such back-projections?
• Which methods are suited to users outside the field? How can the necessity be avoided to set dedicated parameters by hand or choose from different possible algorithms by hand?

Members

• Michael Aupetit, CEA LIST
• Michael Biehl, RU Groningen
• Barbara Hammer, Bielefeld University
• Alexander Hinneburg, Martin Luther University Halle
• Ata Kaban, University of Birmingham
• Samuel Kaski, Aalto University
• Daniel Keim, University of Konstanz
• Marc E. Latoschik, University of Würzburg
• Guy Lebanon, Georgia Institue of Technology
• John A. Lee, Universite Catholique de Louvain
• Laurens van der Maaten, Delft University of Technology
• Erzsebet Merenyi, Rice University
• Olfa Nasraoui, University of Louisville
• Tim Nattkemper, Bielefeld University
• Fabrice Rossi, University Paris I
• Fei Sha, University of Southern California
• Yasufumi Takama, Tokyo Metropolitan University
• Holger Theisel, University of Magdeburg
• Peter Tino, University of Birmingham
• Michel Verleysen, Universite Catholique de Louvain
• Thomas Villmann, University of Applied Sciences Mittweida
• Axel Wismueller, University of Rochester
• Stefan Wrobel, Fraunhofer Institute for Intelligent Analysis and Information Systems
• Gabriel Zachmann, University of Bremen