Methods that find structures in data, such as cluster analysis techniques, are the object of increased interest, spurred by the availability of large datasets accessible though distributed information networks. Since their inception, clustering approaches based on the theory of fuzzy sets have been extensively applied due to their rich representational capabilities, their formal mathematical underpinnings, and the relations between the nature of fuzzy classifications and utilitarian and metric concepts such as preferences and similarities. Furthermore, the very nature of fuzzy clustering methods readily permits the definition of interesting data structures as instances of paradigmatic models that are approximated by subsets of the dataset being analyzed.

In our presentation, we will review the motivation and evolution of fuzzy-set based methods to discover structures in data. Our point of departure will be the initial proposal for the formulation of relational fuzzy clustering as an optimization problem over the set of all partitions of a subset of a metric space. We will also examine the related problem of partitioning a subset of a vector space and discuss major approaches to its treatment. Continuing our retrospective examination of fuzzy-clustering techniques we will focus on significant milestones in the evolution of this methodology including the generalizations of the notions of prototype, clustering, and fuzzy partition.

The objective of this review is to motivate the introduction of qualitative object description methods, which are soft-computing based approaches for the representation of complex objects in terms of significant qualitative features (e.g., interesting substructures in biological molecules) and by identification of qualitative relationships between those features (e.g., spatial relations between features). The purpose of these methods is the description of complex objects in terms—called perceptions by Zadeh—that are meaningful to domain experts and the study of collections of these objects on the basis of these representations.

The ultimate goal of our presentation, however, is the discussion of recent approaches based on the qualitative description of objects in datasets from multiple viewpoints and the joint study of the results of these classification exercises. These methods, which are well suited for implementation in multiagent, distributed, collaborative environments, extract additional knowledge by fusion and comparative evaluation of the data structures discovered employing different variables, models, and viewpoints.

Three components —representation, inference and learning— are critical in constructing an intelligent system. We need a declarative representation that is a reasonable encoding of our world model. We need to be able to use this representation effectively to answer a broad range of questions that are of interest. And we need to be able to acquire this probability distribution, combining expert knowledge and accumulated data. Probabilistic graphical models support all three capabilities for a broad range of problems.

Evolutionary computation has become an essential tool for solving difficult and high-dimensional optimization problems in a broad range of real problems. Genetic algorithms have been the subject of the major part of such applications. Estimation of distribution algorithms offer a recent evolutionary paradigm that constitutes a natural and attractive alternative to genetic algorithms. They make use of a probabilistic model, learnt from the promising solutions, to guide the search process.

This talk will review synergies between probabilistic graphical models and evolutionary computation. First, we will show how to use evolutionary computation in inference and in learning from data problems within probabilistic graphical models. The search for the maximum a posteriori assignment and the optimal triangulation of the moral graph will exemplify inference problems. Learning from data may be carried out both in the space of directed acyclic graphs and in the space of orderings. Second, we will illustrate how to use Bayesian networks and Gaussian networks for developing estimation of distribution algorithms in discrete and continuous domains, respectively. Third, recent advances will be presented, covering regularization methods for learning probabilistic graphical models from data, multi-label classification with multidimensional Bayesian networks classifiers and estimation of distribution algorithms based on copulas and Markov networks. The talk will finish with some challenging applications in bioinformatics and neuroscience.

In many application areas data is being collected over a long time. Due to its temporal nature such data not only captures influences, like market forces or the launch of a competing product, but also reflects the changes of the underlying domain. Often, change can mean a risk or an opportunity. In either case, it is in many domains not only imperative to detect change in order to survive or to win but also it is inevitable for successful decision making to analyze and act upon it. For example, churning customer groups must be identified as such before they comprise a considerable portion of the target group. Quality issues associated with a vehicle manufacturer have to be treated before affecting thousands of units.

As a response to these requirements there is an increasing research interest in methods that aim at analyzing the changes within a domain. Two different research areas can be distinguished. The area of temporal data mining aims at finding patterns which repeat in time, whereas change mining aims to describe how the results of data mining, models and patterns, evolve over time. While being considerably different in their objectives both areas have in common that they often employ frequent pattern mining approaches, either to express or to analyze changes.

In this talk I will contrast both fields from an application point of view on the example of frequent pattern mining. Within the area of temporal data mining I will discuss how association rules can be modified to allow for constraining temporal relations such as "occurred before" or "happen while". Within the area of change mining I will discuss how frequent patterns can be analyzed for their change characteristics using statistical, template-based and visual approaches. I will give examples of successful industrial applications dealing with different models of incorporating the temporal aspects. More specific, applications with automobile manufacturers and telecommunication providers will be presented and provide evidence of the interest and needs of such analysis methods.

To understand the details of the brain function, a large scale system model that reflects structure and neurophysiological characteristics needs to be implemented. Though numerous computational models of different brain areas have been proposed, its integration for the development of a large scale model have not yet been accomplished because these models were described by different programming languages and mostly because they use different data formats. We introduced a platform for a collaborative brain system modeling (PLATO) where one can construct computational models using several programming languages and connect them at the I/O level with a common data format namely netCDF. As an example, a whole visual system model including eye movement, eye optics, retinal network and visual cortex is being developed. Preliminary results demonstrate that the integrated model successfully simulates the signal processing flow at different stages of the visual system.

Key words: Neuroinformatics, Large scale modeling, Visual system, Common data format

Rough sets were introduced by Z. Pawlak in 1982 as a formal approximation of a crisp set in terms of a pair of sets, which provide the lower and the upper approximation of the original set. Rough sets provide an important and mathematically established tool for dimensionality reduction in large data. A lot of research has been undertaken, over the last decade, for integrating rough sets into the broader framework of computational intelligence. I will try to provide an overview of such hybridization, along with some applications.

The talk begins with a brief introduction to rough sets. This will be followed by describing some of the hybridizations of rough sets with neural networks, fuzzy sets and genetic algorithms, in well-known tasks of pattern recognition and data mining. Applications are presented for knowledge encoding, rule extraction, dimensionality reduction and clustering. Results demonstrate the suitability of the methodology for feature selection with improved recognition, in diverse domains like microarray gene expressions and face recognition.