Rough Sets and Current Trends in Computing

Large scale networked systems that include heterogeneous entities, e.g., humans and computational entities are becoming increasingly prevalent. Prominent applications include the Internet, large scale disaster relief and network centric warfare. In such systems, large heterogeneous coordinating entities exchange uncertain information to obtain situation awareness. Uncertain and possibly conflicting sensor data is shared across a peer-to-peer network. Not every team member will have direct access to sensors and team members will be influenced mostly by their neighbors in the network with whom they communicate directly. In this talk I will present our work on the dynamics and emergent behaviors of a large team sharing beliefs to reach conclusions about the world. Unlike past work, the nodes in the networks we study are autonomous and actively fuse information they receive. Nodes can change their beliefs as they receive additional information over time.

Dominance-based Rough Set Approach (DRSA) has been proposed as an extension of the Pawlak’s concept of Rough Sets in order to deal with ordinal data [see [2,3]]. Ordinal data are typically encountered in multi-attribute decision problems where a set of objects (also called actions, acts, solutions, etc.) evaluated by a set of attributes (also called criteria, variables, features, etc.) raises one of the following questions: (i) how to assign the objects to some ordered classes (ordinal classification), (ii) how to choose the best subset of objects (optimization), or (iii) how to rank the objects from the best to the worst (ranking). The answer to everyone of these questions involves an aggregation of the multi-attribute evaluation of objects, which takes into account a law relating the evaluation and the classification, or optimization, or ranking decision. This law has to be discovered from the data by inductive learning. In case of decision problems corresponding to some physical phenomena, this law is a model of cause-effect relationships, and in case of a human decision making, this law is a decision maker’s preference model. In DRSA, these models have the form of a set of “

if..., then...

decision rules. In case of multi-attribute classification the syntax of rules is: “

if evaluation of object a is better (or worse) than given values of some attributes, then a belongs to at least (at most) given class”

, and in case of multi-attribute optimization or ranking: “

if object a is preferred to object b in at least (at most) given degrees with respect to some attributes, then a is preferred to b in at least (at most) given degree”

.

RSCTC’2010 Discovery Challenge was a special event of Rough Sets and Current Trends in Computing conference. The challenge was organized in the form of an interactive on-line competition, at

TunedIT

.org platform, in days between Dec 1, 2009 and Feb 28, 2010. The task was related to feature selection in analysis of DNA microarray data and classification of samples for the purpose of medical diagnosis or treatment. Prizes were awarded to the best solutions. This paper describes organization of the competition and the winning solutions.

In this paper we present

TunedIT

system which facilitates evaluation and comparison of machine-learning algorithms.

TunedIT

is composed of three complementary and interconnected components: TunedTester, Repository and Knowledge Base.

TunedTester is a stand-alone Java application that runs automated tests (experiments) of algorithms. Repository is a database of algorithms, datasets and evaluation procedures used by TunedTester for setting up a test. Knowledge Base is a database of test results. Repository and Knowledge Base are accessible through

TunedIT

website.

TunedIT

is open and free for use by any researcher. Every registered user can upload new resources to Repository, run experiments with TunedTester, send results to Knowledge Base and browse all collected results, generated either by himself or by others.

As a special functionality, built upon the framework of automated tests,

TunedIT

provides a platform for organization of on-line interactive competitions for machine-learning problems. This functionality may be used, for instance, by teachers to launch contests for their students instead of traditional assignment tasks; or by organizers of machine-learning and data-mining conferences to launch competitions for the scientific community, in association with the conference.

In this article, an evolutionary crisp clustering technique is described that uses a new consensus multiobjective differential evolution. The algorithm is therefore able to optimize two conflicting cluster validity measures simultaneously and provides resultant Pareto optimal set of non-dominated solutions. Thereafter the problem of choosing the best solution from resultant Pareto optimal set is resolved by creation of consensus clusters using voting procedure. The proposed method is used for analyzing the categorical data where no such natural ordering can be found among the elements in categorical domain. Hence no inherent distance measure, like the Euclidean distance, would work to compute the distance between two categorical objects. Index-coded encoding of the cluster medoids (centres) is used for this purpose. The effectiveness of the proposed technique is provided for artificial and real life categorical data sets. Also statistical significance test has been carried out to establish the statistical significance of the clustering results. Matlab version of the software is available at http://bio.icm.edu.pl/~darman/CMODECC.

In numerous data clustering problems, the main priority remains a constant demand on development of new improved algorithmic schemes capable of robust and correct data handling. This requirement has been recently boosted by emerging new technologies in data acquisition area. In image processing and image analysis procedures, the image segmentation procedures have the most important impact on the image analysis results.

In data analysis methods, in order to improve understanding and description of data structures, many innovative approaches have been introduced. Data analysis methods always strongly depend upon revealing inherent data structure. In the paper, a new algorithmic

R

ough

E

ntropy

F

ramework - (REF, in short) has been applied in the probabilistic setting. Crisp and Fuzzy RECA measures (

R

ough

E

ntropy

C

lustering

A

lgorithm) introduced in [5] are extended into probability area. The basic rough entropy notions, the procedure of rough (entropy) measure calculations and examples of probabilistic approximations have been presented and supported by comparison to crisp and fuzzy rough entropy measures. In this way, uncertainty measures have been combined with probabilistic procedures in order to obtain better insight into data internal structure.

Average-link (AL) is a distance based hierarchical clustering method, which is not sensitive to the noisy patterns. However, like all hierarchical clustering methods AL also needs to scan the dataset many times. AL has time and space complexity of

O

(

n

2

), where

n

is the size of the dataset. These prohibit the use of AL for large datasets. In this paper, we have proposed a distance based hierarchical clustering method termed

l

-AL which speeds up the classical AL method in any metric (vector or non-vector) space. In this scheme, first leaders clustering method is applied to the dataset to derive a set of leaders and subsequently AL clustering is applied to the leaders. To speed-up the leaders clustering method, reduction in distance computations is also proposed in this paper. Experimental results confirm that the

l

-AL method is considerably faster than the classical AL method yet keeping clustering results at par with the classical AL method.

Grouping data into meaningful clusters is an important data mining task. DBSCAN is recognized as a high quality density-based algorithm for clustering data. It enables both the determination of clusters of any shape and the identification of noise in data. The most time-consuming operation in DBSCAN is the calculation of a neighborhood for each data point. In order to speed up this operation in DBSCAN, the neighborhood calculation is expected to be supported by spatial access methods. DBSCAN, nevertheless, is not efficient in the case of high dimensional data. In this paper, we propose a new efficient TI DBSCAN algorithm and its variant TI-DBSCAN-REF that apply the same clustering methodology as DBSCAN. Unlike DBSCAN, TI-DBSCAN and TI-DBSCAN-REF do not use spatial indices; instead they use the triangle inequality property to quickly reduce the neighborhood search space. The experimental results prove that the new algorithms are up to three orders of magnitude faster than DBSCAN, and efficiently cluster both low and high dimensional data.

Experiments and results regarding vehicle type classification are presented. Three classes of vehicles are recognized: sedans, vans and trucks. The system uses a non-calibrated traffic camera, therefore no direct vehicle dimensions are used. Various vehicle descriptors are tested, including those based on vehicle mask only and those based on vehicle images. The latter ones employ Speeded Up Robust Features (SURF) and gradient images convolved with Gabor filters. Vehicle type is recognized with various classifiers: artificial neural network, K-nearest neighbors algorithm, decision tree and random forest.

Results of experiments regarding lip gesture recognition with an artificial neural network are discussed. The neural network module forms the core element of a multimodal human-computer interface called LipMouse. This solution allows a user to work on a computer using lip movements and gestures. A user face is detected in a video stream from a standard web camera using a cascade of boosted classifiers working with Haar-like features. Lip region extraction is based on a lip shape approximation calculated by the means of lip image segmentation using fuzzy clustering. ANN is fed with a feature vector describing lip region appearance. The descriptors used include a luminance histogram, statistical moments and co-occurrence matrices statistical parameters. ANN is able to recognize with a good accuracy three lip gestures: mouth opening, sticking out the tongue and forming puckered lips.

A fuzzy logic inference system was created, based on the analysis of animated motion features. The objective of the system is to facilitate the creation of high quality animation by analyzing personalized styles contained in numerous animations. Sequences portraying a virtual character acting with a differentiating personalized style (natural or exaggerated) and various levels of fluidity were prepared and subjectively evaluated. Knowledge gathered in subjective evaluation tests was processed utilizing variable precision rough set (VPRS) approach for defining non-ambiguous inverse relation between subjective features of the result animation and objective parameters of the animated motion. Once the mapping is known then the user can define own requirements on animation, and the input motion is processed accordingly to produce the desired result. The paper focuses on employing variable precision rough set methodology for selection of representative parameter values.

The current work describes a phoneme matching algorithm based on rough set concepts. The objective of this type of algorithms is focused on the localization of the phonemic content of a specific spoken occurrence. According to the proposed algorithm, a number of rough sets containing the multiple expected phonemic instances in a sequence are created, each defined by a set of short term frames of the voice signal. The properties of the corresponding information system are derived from a features set calculated from the speech signal upon initiation. Given the above, an iterative procedure is applied by updating the phoneme instances versus the optimization of the accuracy metric. The main advantage of this algorithm is the absence of a training phase allowing for wider speaker adaptability and independency. The current paper focuses on the feasibility of the task as this work is still in early research stage.

The paper presents how rule-based processing can be applied to automatically evaluate the motor state of Parkinson’s Disease patients. Automatic monitoring of patients by using biometric sensors can provide assessment of the Parkinson’s Disease symptoms. All data on PD patients’ state are compared to historical data stored in the database and then a rule-based decision is applied to assess the overall illness state. The training procedure based on doctors’ questionnaires is presented. These data constitute the input of several rule-based classifiers. It has been proved that the rough-set-based algorithm can be very suitable for automatic assessment of the PD patient’s stability/worsening state.

A large amount of digital video data is stored in local or network visual retrieval systems. The new technology advances in multimedia information processing as well as in network transmission have made video data publicly and relatively easy available. Users need the adequate tools to locate their desired video or video segments quickly and efficiently, for example in Internet video collections, TV shows archives, video-on-demand systems, personal video archives offered by many public Internet services, etc. Detection of scenes in TV videos is difficult because the diversity of effects used in video editing puts up a barrier to construct an appropriate model. The framework of automatic recognition and classification of scenes reporting the sport events in a given discipline in TV sports news have been proposed. Experimental results show good performance of the proposed scheme on detecting scenes on a given sport discipline in TV sports news. In the tests a special software called AVI – the Automatic Video Indexer has been used to detect shots and then scenes in tested TV news videos.

Blending is a well-established technique, commonly used to increase performance of predictive models. Its effectiveness has been confirmed in practice as most of the latest international data-mining contest winners were using some kind of a committee of classifiers to produce their final entry. This paper is a technical report presenting a method of using a genetic algorithm to optimize an ensemble of multiple classification or regression models. An implementation of this method in

R

, called Genetic Meta-Blender, was tested during the Australian Data Mining 2009 Analytic Challenge competition and it was awarded with the Grand Champion prize for achieving the best overall result. In the report, the purpose of the challenge is described and details of the winning approach are given. The results of Genetic Meta-Blender are also discussed and compared to several baseline scores.

Argument based learning allows experts to express their domain, local knowledge about the circumstances of making classification decisions for some learning examples. In this paper we have incorporated this idea in rule induction as a generalization of the MODLEM algorithm. To adjust the algorithm to the redefined task, a new measure for evaluating rule conditions and a new classification strategy with rules had to be introduced. Experimental studies showed that using arguments improved classification accuracy and structure of rules. Moreover the proper argumentation improved recognition of the minority class in imbalanced data without essential decreasing recognition of the majority classes.

In the paper we present a new framework for improving classifiers learned from imbalanced data. This framework integrates the SPIDER method for selective data pre-processing with the Ivotes ensemble. The goal of such integration is to obtain improved balance between the sensitivity and specificity for the minority class in comparison to a single classifier combined with SPIDER, and to keep overall accuracy on a similar level. The IIvotes framework was evaluated in a series of experiments, in which we tested its performance with two types of component classifiers (tree- and rule-based). The results show that IIvotes improves evaluation measures. They demonstrated advantages of the abstaining mechanism (i.e., refraining from predictions by component classifiers) in IIvotes rule ensembles.

In this paper we studied re-sampling methods for learning classifiers from imbalanced data. We carried out a series of experiments on artificial data sets to explore the impact of noisy and borderline examples from the minority class on the classifier performance. Results showed that if data was sufficiently disturbed by these factors, then the focused re-sampling methods – NCR and our SPIDER2 – strongly outperformed the oversampling methods. They were also better for real-life data, where PCA visualizations suggested possible existence of noisy examples and large overlapping ares between classes.

The problem of recurring concepts in data stream classification is a special case of concept drift where concepts may reappear. Although several methods have been proposed that are able to learn in the presence of concept drift, few consider concept recurrence and integration of context. In this work, we extend existing drift detection methods to deal with this problem by exploiting context information associated with learned decision models in situations where concepts reappear. The preliminary experimental results demonstrate the effectiveness of the proposed approach for data stream classification problems with recurring concepts.

Support Feature Machines (SFM) define useful features derived from similarity to support vectors (kernel transformations), global projections (linear or perceptron-style) and localized projections. Explicit construction of extended feature spaces enables control over selection of features, complexity control and allows final analysis by any classification method. Additionally projections of high-dimensional data may be used to estimate and display confidence of predictions. This approach has been applied to the DNA microarray data.

We propose a framework for experimental verification whether mechanisms of voting among rough-set-based classifiers and criteria for extracting those classifiers from data should follow analogous mathematical principles. Moreover, we show that some of types of criteria perform better for high-quality data while the others are useful rather for low-quality data. The framework is based on the principles of approximate attribute reduction and probabilistic extensions of rough-set-based approach to data analysis. The framework is not supposed to produce the best-ever classification results, unless it is extended by some additional parameters known from the literature. Instead, our major goal is to illustrate in a possibly simplistic way that it is worth unifying mathematical background for the stages of learning and applying rough-set-based classifiers.

Co-training is an effective semi-supervised learning method which uses unlabeled instances to improve prediction accuracy. In the co-training process, a random sampling is used to gradually select unlabeled instances to train classifiers. In this paper we explore whether other sampling methods can improve co-training performance. A novel selective sampling method, agreement-based sampling, is proposed. Experimental results show that our new sampling method can improve co-training significantly.

Clinical Decision Support Systems embed data-driven decision models designed to represent clinical acumen of an experienced physician. We argue that eliminating physicians’ diagnostic biases from data improves the overall quality of concepts, which we represent as decision rules. Experiments conducted on prospectively collected clinical data show that analyzing this filtered data produces rules with better coverage, certainty and confirmation. Cross-validation testing shows improvement in classification performance.

To emphasize gene interactions in the classification algorithms, a new representation is proposed, comprising gene-pairs and not single genes. Each pair is represented by

L

1

difference in the corresponding expression values. The novel representation is evaluated on benchmark datasets and is shown to often increase classification accuracy for genetic datasets. Exploiting the gene-pair representation and the Gene Ontology (GO), the semantic similarity of gene pairs can be incorporated to pre-select pairs with a high similarity value. The GO-based feature selection approach is compared to the plain data driven selection and is shown to often increase classification accuracy.

In the paper, several notions of rough satisfiability of formulas are recalled and discussed from the standpoint of fuzzy set theory. By doing so we aim to better understand what rough satisfiability really means and to look for schemata describing it.

In the paper we consider a topological approximation space (

U

,

τ

) (induced by a given information system

$\mathcal{I}$

) as a discrete dynamical system; that is, we are concerned with a finite approximation space

U

whose topology

τ

is induced by a function

f

:

U

→

U

. Our aim is to characterise these type of approximation spaces by means of orbits which represent the evolution of points of

U

with respect to the process

f

. Apart from topological considerations we also provide some algebraic characterisation of orbits. Due to the finiteness condition imposed by

$\mathcal{I}$

, any point

a

 ∈ 

U

is eventually cyclic. In consequence, as we demonstrate, orbits are algebraically close to rough sets, e.g. they induce a Łukasiewicz algebra of order two, where the lower approximation operator may be interpreted as the action of retriving a cycle from a given orbit and the upper approximation operator may be interpreted as the action of making a given orbit cyclic.

We propose a generalization of Pawlak’s rough set model for the multi-agent situation, where information from an agent can be preferred over that of another agent of the system while deciding membership of objects. Notions of lower/upper approximations are given which depend on the knowledge base of the sources as well as on the position of the sources in the hierarchy giving the preference of sources. Some direct consequences of the definitions are presented.

A classification of the different dynamics which can arise in rough sets is given, starting from three different standpoints: information tables, approximation spaces and coverings. Dynamics is intended in two broad meanings: evolution in time and originated from different sources. Existing works on this topic are then categorized accordingly.

A set of subsets of a set may be seen as granules that allow arbitrary subsets to be approximated in terms of these granules. In the simplest case of rough set theory, the set of granules is required to partition the underlying set, but granulations based on relations more general than equivalence relations are well-known within rough set theory. The operations of dilation and erosion from mathematical morphology, together with their converse forms, can be used to organize different techniques of granular approximation for subsets of a set with respect to an arbitrary relation. The extension of this approach to granulations of sets with structure is examined here for the case of hypergraphs. A novel notion of relation on a hypergraph is presented, and the application of these relations to a theory of granularity for hypergraphs is discussed.

This paper introduces a perceptual tolerance intersection of sets as an example of near set operations. Such operations are motivated by the need to consider similarities between digital images viewed as disjoint sets of points. The proposed approach is in keeping with work by E.C. Zeeman on tolerance spaces and visual perception and J.H. Poincaré on sets of similar sensations used to define representative (aka

tolerance

) spaces such as visual, tactile and motile spaces. Perceptual tolerance intersection of sets is a direct consequence of recent work on near sets and a solution to the problem of how one goes about discovering affinities between digital images. The main contribution of this article is a description-based approach to assessing the resemblances between digital images.

In this paper, we define a category

R-APR

whose objects are sets and morphisms are the pairs of rough set approximation operators. We show that

R-APR

is isomorphic to a full subcategory of the category

cdrTex

whose objects are complemented textures and morphisms are complemented direlations. Therefore,

cdrTex

may be regarded as an abstract model for the study of rough set theory. On the other hand, dagger symmetric monoidal categories play a central role in the abstract quantum mechanics. Here, we show that

R-APR

and

cdrTex

are also dagger symmetric monoidal categories.

We discuss some important issues for applications that are related to generalizations of the 1994 approximation space definition [11]. In particular, we present examples of rough set based strategies for extension of approximation spaces from samples of objects onto the whole universe of objects. This makes it possible to present methods for inducing approximations of concepts or classifications analogously to the approaches for inducing classifiers known in machine learning or data mining.

The paper is devoted to the formalization of two elementary but important problems within rough set theory. We mean searching for the minimal requirements of the well-known rough operators – the lower and the upper approximations in an abstract approximation space to retain their natural properties. We also discuss pros and cons of the development of the computer-checked repository for rough set theory based on the comparison of certain rough approximation operators proposed by Anna Gomolińska.

Communicative approximations, as used in language, are equivalence relations that partition a continuum, as opposed to observational approximations on the continuum. While the latter can be addressed using tolerance interval approximations on interval algebra, new constructs are necessary for considering the former, including the notion of a “rough interval”, which is the indiscernibility region for an event described in language, and “rough points” for quantities and moments. We develop the set of qualitative relations for points and intervals in this “communicative approximation space”, and relate them to existing relations in exact and tolerance-interval formalisms. We also discuss the nature of the resulting algebra.

There is a natural generalization of an indiscernibility relation used in rough set theory, where rather than partitioning the universe of discourse into indiscernibility classes, one can consider a covering of the universe by similarity-based neighborhoods with lower and upper approximations of relations defined via the neighborhoods. When taking this step, there is a need to tune approximate reasoning to the desired accuracy. We provide a framework for analyzing self-adaptive knowledge structures. We focus on studying the interaction between inputs and output concepts in approximate reasoning. The problems we address are:

given similarity relations modeling approximate concepts, what are similarity relations for the output concepts that guarantee correctness of reasoning?

assuming that output similarity relations lead to concepts which are not accurate enough, how can one tune input similarities?

Unit operations are some special functions on sets. The concept of the unit operation originates from researches of U.Wybraniec-Skardowska. The paper is concerned with the general properties of such functions. The isomorphism between binary relations and unit operations is proved. Algebraic structures of families of unit operations corresponding to certain classes of binary relations are considered. Unit operations are useful in Pawlak’s Rough Set Theory. It is shown that unit operations are upper approximations in approximation space. We prove, that in the approximation space (

U

,

R

) generated by a reflexive relation

R

the corresponding unit operation is the least definable approximation if and only if the relation

R

is transitive.

The nearest neighbor classification is a simple and effective technique for pattern recognition. The performance of this technique is known to be sensitive to the distance function used in classifying a test instance. In this paper, we propose a technique to learn sample weights via maximizing classification consistency. Experimental analysis shows that the distance trained in this way enlarges the classification consistency on several datasets and has a strong ability to tolerate noise. Moreover, the proposed approach has better performance than nearest neighbor classification and several state-of-the-art methods.

The article reports our implementation of a rough set-based incremental learning algorithm involving the application of the hierarchy of probabilistic decision tables to face recognition. The implementation, the related theoretical background such as the basics of the variable precision rough set theory, the algorithm, the classifier structure and experiments with balanced and imbalanced data sets are presented.

In this paper, we propose a new approach of classification based on rough sets denoted Dynamic Belief Rough Set Classifier (D-BRSC) which is able to learn decision rules from uncertain data. The uncertainty appears only in decision attributes and is handled by the Transferable Belief Model (TBM), one interpretation of the belief function theory. The feature selection step of the construction procedure of our new technique of classification is based on the calculation of dynamic reduct. The reduction of uncertain and noisy decision table using dynamic approach which extracts more relevant and stable features yields more significant decision rules for the classification of the unseen objects. To prove that, we carry experimentations on real databases using the classification accuracy criterion. We also compare the results of D-BRSC with those obtained from Static Belief Rough Set Classifier (S-BRSC).

Non

-

deterministic

Information

Systems

(

NISs

) are well known as systems for handling information incompleteness in data. In our previous work, we have proposed

NIS

-

Apriori

algorithm aimed at extraction of decision rules from

NISs

.

NIS

-

Apriori

employs the minimum and the maximum supports for each descriptor, and it effectively calculates the criterion values for defining rules. In this paper, we focus on

$Lipski\mbox{'}s$

Incomplete

Information

Databases

(

IIDs

), which handle non-deterministic information by means of the sets of values and intervals. We clarify how to understand decision rules in

IIDs

and appropriately adapt our

NIS

-

Apriori

algorithm to generate them. Rule generation in

IIDs

turns out to be more flexible than in

NISs

.

Randomisation is a method to test the statistical significance of a symbolic rule; it is, however, very expensive. In this paper we present a sequential randomisation test which in most cases greatly reduces the number of steps needed for a conclusion.

We propose an ensemble method that solves ordinal classification problem with monotonicity constraints. The classification data is structured using the Variable Consistency Dominance-based Rough Set Approach (VC-DRSA). The method employs a variable consistency bagging scheme to produce bootstrap samples that privilege objects (i.e., classification examples) with relatively high values of consistency measures used in VC-DRSA. In result, one obtains an ensemble of rule classifiers learned on bootstrap samples. Due to diversification of bootstrap samples controlled by consistency measures, the ensemble of classifiers gets more accurate, which has been acknowledged by a computational experiment on benchmark data.

We consider learning abilities of classifiers learned from data structured by rough set approaches into lower approximations of considered sets of objects. We introduce two measures,

λ

and

δ

, that estimate attainable predictive accuracy of rough-set-based classifiers. To check the usefulness of the estimates for various types of classifiers, we perform a computational experiment on fourteen data sets. In the experiment, we use two versions of the rough-set-based rule classifier, called VC-DomLEM, and few other well known classifiers. The results show that both introduced measures are useful for an a priori identification of data sets that are hard to learn by all classifiers.

In the paper, we consider the notions of exact and approximate decision reducts for binary decision tables. We present upper bounds on minimum cardinality of exact and approximate reducts depending on the number of rows (objects) in the decision table. We show that the bound for exact reducts is unimprovable in the general case, and the bound for approximate reducts is almost unimprovable in the general case.

This paper concerns some aspects of mathematical flow graph based data analysis. In particular, taking a flow graph view on rough sets’ categories and measures leads to a new methodology of inductively reasoning form data. This perspective shows interesting relationships and properties among rough set, flow graphs and inverse flow graphs. A possible car dealer application is outlined and discussed. Evidently, our new categories and measures assist and alleviate some limitations in flow graphs to discover new patterns and explanations.

ARES System was a data analysis tool supporting Rough Set theory. It has been expanded to cover other approaches like Emerging Patterns and Support Vector Machine. A special feature of ARES System is ability to identify exceptional objects within information systems by using credibility coefficients. The credibility coefficient is a measure, which attempts to weigh up a degree of typicality of each object in respect to the rest of information system. The paper presents an idea of credibility coefficients based on SVM approach. The new coefficients are compared with the others ones available in the ARES System.

The paper describes an algorithm that constructs approximate decision trees (

α

-decision trees), which are optimal relatively to one of the following complexity measures: depth, total path length or number of nodes. The algorithm uses dynamic programming and extends methods described in [4] to constructing approximate decision trees. Adjustable approximation rate allows controlling algorithm complexity. The algorithm is applied to build optimal

α

-decision trees for two data sets from UCI Machine Learning Repository [1].

Domain knowledge elicitation constitutes a crucial task in designing effective machine learning algorithm, and is often indispensable in problem domains that display a high degree of internal complexity such as knowledge discovery and data mining, the recognition of structured objects, human behavior prediction, or multi-agent cooperation. We show how to facilitate this difficult and sometimes tedious task with a hierarchical concept learning scheme, designed to cope with the inherent vagueness and complexity of knowledge therein used. We present how our approach, based on Rough Mereology and Approximate Reasoning frameworks, correlate to other well established approaches to machine learning.

Mobile agents are small programs that can transport their code, data and execution context from one machine to another and be capable of continuing execution in the new environment. This technology has a lot of advantages and promising applications, unfortunately there is a noticeable absence of deployed solutions. There are few reasons of this situation but one of the most important is the lack of tools that can be used for the configuration management of mobile agents. This process focuses on the monitoring and controlling configuration items and is essential for other processes like incident management or availability management. In this paper a new, flexible and universal solution for the configuration management of mobile agents is proposed. This solution is based on well known and widely used management standard -

SNMP

(Simple Network Management Protocol).

An adaptive immune-inspired multiagent system (AIBMAS) is proposed. The intelligence behind such system is based on the idiotypic immune network. Tunable activation Threshold (TAT) proposes that agents adapt their activation thresholds. Immune algorithm based on the immune network theory and memory mechanism is derived.

The paper describes how Distributed Default Logic (DDL) can be used as a formalism for context-aware computing in a Multi-Agent System. It is shown that the original notation does not require any changes. The DDL reasoning engine has been adapted to handle situations like unavailability of sensors. New semantics of Distributed Default Rules in the application to reasoning with context information is also described.

In a multi-agent system the sought information can often be found across various knowledge bases, which means that making early assumptions can lead to hasty conclusions. In the paper we present a formalism for distributed default reasoning to be performed by a group of agents that share knowledge in the form of a distributed default theory. The formalism is based on default transformations, which can be used to derive answers to queries in the form of defaults. Such new defaults can then be treated as intermediate results in the reasoning process. It is shown that passing messages containing transformed defaults is more informative than strict statements and enables avoiding early conclusions. Moreover, the extended reasoning features are embedded in the description logic framework.

This paper describes a platform for testing automated argumentation strategies for agents. It is a continuation of the discussion about the Arguing Agents Competition (AAC). The second version of the AAC platform is introduced, including the architecture and the capabilities of the platform, the currently available engine and an automated strategy for the dialogue game.

The argumentation and some of the formalizations for the arguments and dialogues are briefly presented.

Many kinds of attributes are used for various areas of decision making. Sometimes the attributes have complicated vector-types as in MPEG-7 visual descriptors that prevent us from attaching unequal importance to each descriptor for the construction of content- or emotion-based image retrievals. In this paper, fuzzy similarity-based rough approximation is used for determining the relative importance of MPEG-7 visual descriptors for an emotion. In the methods, the relative importance is given to a descriptor itself rather than a component of the vector of a descriptor or a combined descriptor. Also we propose a method for building a classification system based on representative color images. The experimental result shows the proposed classification method is promising for the emotional classification or evaluation of color images.

Microblogging is a social media tool that allows users to write short text messages to public and private networks. This research focuses specifically on the microblogging on Facebook. The main purposes of this study are to investigate and compare what recommendation sources influence the intention to use microbloggings and to combine gender, daily internet hour usage and past use experience to infer the usage of microbloggings decision rules using a dominance-based rough-set approach (DRSA). Data for this study were collected from 382 users and potential users. The analysis is grounded in the taxonomy of induction-related activities using DRSA to infer the usage of microbloggings decision rules. Finally, the study of the nature of microblogging reflects essential practical and academic value.

The paper describes the concept and the architecture of the data warehouse and reporting modules dedicated to distributed fault injection testbench. The purpose of this warehouse is to collect the data from fault simulation experiments and support researchers in exploration and analysis of these results. The data model of the warehouse, main ETL processes, multidimensional structure of OLAP cube, and predefined reports are discussed. Practical advantages of the presented system are illustrated with some exemplary analyses of the experimental results collected during dependability evaluation of the chemical reactor control algorithm with software implemented fault injection approach.

The Location Management problem is an important issue of mobility management, which is responsible for determining the network configuration, with the major goal of minimizing the involved costs. One of the most common strategies of location management is the Reporting Cells (RC) scheme, which mainly considers the location update and the paging costs. In this paper we propose a Scatter Search (SS) based approach applied to the Reporting Cells as a cost optimizing solution, with the objective of achieving the best network configuration defining a subset of cells as reporting cells and the others as non-reporting cells. With this work we want to define the most adequate values of the SS parameters, when applied to the RC problem, using twelve test networks that represent 4 distinct groups divided by size. We also want to compare the performance of this SS based approach with a previous study based on Differential Evolution and also with other approaches presented in the literature. The results obtained are very interesting because they outperform those obtained with other approaches exposed in the literature.

This work attempts to report the stylistic differences in blogging for gender and age group variations using slang word co-occurrences. We have mainly focused on co-occurrence of non dictionary words across bloggers of different gender and age groups. For this analysis, we have focused on the feature

use of slang words

to study the stylistic variations of bloggers across various age groups and gender. We have modeled the co-occurrences of slang words used by bloggers as graph based model where nodes are slang words and edges represent the number of cooccurrences and studied the variations in predicting age groups and gender. We have used demographically tagged blog corpus from ICWSM Spinner dataset for these experiments and used Naive Bayes classifier with 10 fold cross validations. Preliminary results shows that the concurrence of of slang words could be a better choice for predicting age and gender.

The easily reconfigurable predictive controllers are supplemented with a mechanism of disturbance measurement utilization. It is done in such a way that the main advantage of the controllers – their simplicity – is maintained. The predictive controllers under consideration are based on fuzzy Takagi–Sugeno (TS) models in which step responses are used as local models. These models are supplemented with the parts describing the influence of disturbances on the outputs of the control plant. Then, the controllers are formulated in such a way that the control signals are easily generated. Efficiency and usefulness of the predictive controllers utilizing disturbance measurement is demonstrated in the example control system of a nonlinear control plant with delay.

In this paper we have proposed a biometric-based authentication system based on rough set theory. The system employed signature for authentication purpose. The major functional blocks of the proposed system are presented. Information is extracted as time functions of various dynamic properties of the signatures. We apply our methodology to global features extracted from a 108-users database. Thirty-one features were identified and extracted from each signature. Rough set approach has resulted in a reduced set of nine features that were found to capture the essential characteristics required for signature identification. Low error rates obtained in experiments illustrate the feasibility of using Rough Set as a promising technique for online signature identification systems.

Eyewitness identification remains an important element in judicial proceedings. It is very convincing, yet it is not very accurate. To better understand eyewitness identification, we began by examining how people understand similarity. This paper reports on analysis of study that examined how people made similarity judgements amongst a variety of facial photographs: participants were presented with a randomly ordered set of photos, with equal numbers of Caucasian (C) and First Nations (F), which they sorted based on their individual assessment of similarity. The number of piles made by the participants was not restricted. After sorting was complete, each participant was then asked to label each pile with a description of the pile’s contents. Following the results of an earlier study, we hypothesize that individuals may be using different strategies to assess similarity between photos. In this analysis, we attempt to use the descriptive pile labels (in particular, related to lips and ears) as a means to uncover differences in strategies for which a classifier can be built, using the rough set attribute reduction methodology. In particular, we aim to identify those pairs of photographs that may be the key for verifying an individual’s abilities and strategies when recognizing faces. The paper describes the method for data processing that enabled the comparisons based on labels. Continued success with the same technique as previously reported to filter pairs before performing the rough sets analysis, lends credibility to its use as a general method. The rough set techniques enable the identification of the sets of photograph pairs that are key to the divisions based on various strategies. This may lead to a practical test for people’s abilities, as well as to inferring what discriminations people use in face recognition.

In this paper we investigate the problem of recognizing the full set of instruments playing in a sound mix. Random mixes of 2-5 instruments (out of 14) were created and parameterized to obtain experimental data. Sound samples were taken from 3 audio data sets. For classification purposes, we used a battery of one-instrument sensitive random forest classifiers, and obtained quite good results.

In this paper we present results of an experimental comparison (in terms of an error rate) of rule sets induced by the LERS data mining system with rule sets induced using the probabilistic rough classification (PRC). As follows from our experiments, the performance of LERS (possible rules) is significantly better than the best rule sets induced by PRC with any threshold (two-tailed test, 5% significance level). Additionally, the LERS possible rule approach to rule induction is significantly better than the LERS certain rule approach (two-tailed test, 5% significance level).

When the indiscernibility relation, fundamental to Classical Rough Set Approach, is substituted with dominance relation, it results in Dominance-Based Rough Set Approach to data analysis. It enables support not only for nominal classification tasks, but also when ordinal properties on attribute values can be observed [1], making DRSA methodology well suited for stylometric processing of texts. Stylometry involves handling quantitative features of texts leading to characterisation of authors to the point of recognition of their individual writing styles. As always, selection of attributes is crucial to classification accuracy, as is the construction of a decision algorithm. When minimal cover gives unsatisfactory results, and all rules on examples algorithm returns very high number of rules, usually constraints are imposed by selection of some reduct and limiting the decision algorithm by including within it only rules with certain support. However, reducts are typically numerous and within them some of conditional attributes are used more often than others, which is also true for conditions specified by decision rules. The paper presents observations how the frequency of usage for features reflects on the performance of decision algorithms resulting from selection of rules with conditional attributes exploited most and least often.

Automatic indexing of music instruments for multi-timbre sounds is challenging, especially when partials from different sources are overlapping with each other. Temporal features, which have been successfully applied in monophonic sound timbre identification, failed to isolate music instrument in multi-timbre objects, since the detection of the start and end position of each music segment unit is very difficult. Spectral features of MPEG7 and other popular features provide economic computation but contain limited information about timbre. Being compared to the spectral features, spectrum signature features have less information loss; therefore may identify sound sources in multi-timbre music objects with higher accuracy. However, the high dimensionality of spectrum signature feature set requires intensive computing and causes estimation efficiency problem. To overcome these problems, the authors developed a new multi-resolution system with an iterative spectrum band matching device to provide fast and accurate recognition.

Understanding cellular properties of neurons is central in neuroscience. It is especially important in light of recent discoveries suggesting that similar neural activity can be produced by cells with quite disparate characteristics. Unfortunately, due to experimental constraints, analyzing how the activity of neurons depends on cellular parameters is difficult. Computational modeling of biological neurons allows for exploration of many parameter combinations, without the necessity of a large number of “wet” experiments. However, analysis and interpretation of often very extensive databases of models can be hard. Thus there is a need for efficient algorithms capable of mining such data. This article proposes a rough sets-based approach to the problem of classifying functional and non-functional neuronal models. In addition to presenting a successful application of the theory of rough sets in the field of computational neuroscience, we are hoping to foster with this paper a greater interest among the members of the rough sets community to explore the plethora of important problems in that field.

We discuss various ideas how to implement execution of approximate SQL statements within Infobright database engine. We first outline the engine’s architecture, which is designed entirely to work with standard SQL. We then discuss several possible extensions towards approximate querying and point out at some analogies with the principles of the theory of rough sets. Finally, we present the results of experiments obtained at the prototype level, both with respect to the speed of query execution and the accuracy of approximate answers.

The tertiary structure of a protein molecule is the main factor which determines its function. All information required for a protein to be folded in its natural structure, is coded in its amino acid sequence. The way this sequence folds in the 3D space can be used for determining its function. With the technology innovations, the number of determined protein structures increases every day, so improving the efficiency of protein structure retrieval and classification methods becomes an important research issue. In this paper, we propose a novel protein classifier. Our classifier considers the conformation of protein structure in the 3D space. Namely, our voxel based protein descriptor is used for representing the protein structures. Then, the Support Vector Machine method (SVM) is used for classifying protein structures. The results show that our classifier achieves 78.83% accuracy, while it is faster than other algorithms with comparable accuracy.

This paper describes a nonlinear Model Predictive Control (MPC) algorithm based on neural models. Two neural models are used on-line: from a dynamic model the free trajectory (the influence of the past) is determined, the second neural network approximates the time-varying feedback law. In consequence, the algorithm is characterised by very low computational complexity because the control signal is calculated explicitly, without any on-line optimisation. Moreover, unlike other suboptimal MPC approaches, the necessity of model linearisation and matrix inversion is eliminated. The presented algorithm is compared with linearisation-based MPC and MPC with full nonlinear optimisation in terms of accuracy and computational complexity.

In this paper, a numerical method of solving the inverse heat conduction problem based on the respectively new tool for combinational optimization, named the Artificial Bee Colony algorithm (ABC), is presented. In the first step, the direct heat conduction problem, associated to the considered inverse heat conduction problem, is solved by using the finite difference method. In the second step, the proper functional, based on the least squares method, is minimized by using the ABC algorithm, giving the solution of the considered problem. An example illustrating the precision and effectiveness of the method is also shown. The proposed approach is original and promising.

Efficient Model Predictive Control (MPC) algorithms based on fuzzy Wiener models are proposed in the paper. Thanks to the form of the model the prediction of the control plant output can be easily obtained. It is done in such a way that the MPC algorithm is formulated as a numerically efficient quadratic optimization problem. Moreover, inversion of the static process model, used in other approaches, is avoided. Despite its relative simplicity the algorithm offers practically the same performance as the MPC algorithm in which control signals are generated after solving a nonlinear optimization problem and outperforms the MPC algorithm based on a linear model. The efficacy of the proposed approach is demonstrated in the control system of a nonlinear control plant.

The most widely used reputation models assume uniform users’ preference structure. In this paper a new reputation management model is presented. It is focused on aggregation of community wide reputation in situation when agents do not share the same preference structure. The reputation is interpreted as vectors of attributes that represent several reputation evaluation criteria. Outcomes of the criteria are transformed by the utility functions and assigned subjective probabilities so that the subjective expected utility values can be obtained. Subjective expected utilities are further aggregated by the weighted ordered weighted average (WOWA) operator. The expressive power of subjective utilities concept along with the WOWA aggregation technique provides the reputation management system with a capability to model various preference structures. It is shown with an illustrative example.

Parameters that vary monotonically with fault development are useful in condition monitoring, but not easy to find especially for complex systems. A method using fuzzy preference based rough set model and principle component analysis (PCA) is proposed to generate such an indicator. The fuzzy preference based rough set model is employed to evaluate the monotonic trends of features reflecting machinery conditions. PCA is used to condense the informative features and generate an indicator which can represent the development of machine health condition. The effectiveness of the proposed method is tested for damage level detection of an impeller in a slurry pump.

In this paper we present a graph-based optimization method for information diffusion and attack durability in networks using properties of Complex Networks. We show why and how Complex Networks with

Scale Free

and

Small World

features can help optimize the topology of networks or indicate weak or strong elements of the network. We define some efficiency measures of information diffusion and attack durability in networks. Using these measures we formulate multicriteria optimization problem to choose the best network. We show a practical example of using the method based on an analysis of a few social networks.

We introduce a number of paraconsistent semantics, including three-valued and four-valued semantics, for the description logic

$\mathcal{SROIQ}$

, which is the logical foundation of OWL 2. We then study the relationship between the semantics and paraconsistent reasoning in

$\mathcal{SROIQ}$

w.r.t. some of them through a translation into the traditional semantics. We also present a formalization of rough concepts in

$\mathcal{SROIQ}$

.

In this paper we present a method for representing the granularity for asymmetric, non-Euclidean relational data. It firstly builds a set of binary classifications based on the directional similarity from each object. After that, the strength of discrimination knowledge is quantified as the indiscernibility of objects based on the Jaccard similarity coefficients between the classifications. Fine but weak discrimination knowledge supported by the small number of binary classifications is more likely to be coarsened than those supported by the large number of classifications, and coarsening of discrimination knowledge causes the merging of objects. Accoding to this feature, we represent the hierarchical structure of data granules by a dendrogram generated by applying the complete-linkage hierarchical grouping method to the derived indiscernibility. This enables users to change the coarseness of discrimination knowledge and thus to control the size of granules.

In this paper we discuss the importance of information systems in modeling interactive computations performed on (complex) granules and propose a formal approach to interactive computations based on information systems. The basic concepts of information systems and rough sets are interpreted in the framework of interactive computations. We also show that information systems can be used for modeling more advanced forms of interactions such as hierarchical ones. The role of hierarchical interactions is emphasized in modeling interactive computations. Some illustrative examples of interactions used in the hierarchical multimodal classification method as well as in the ACT-R 6.0 system are reported.

Case Based Reasoning(CBR), an artificial intelligence technique, solves new problem by reusing solutions of previously solved similar cases. In conventional CBR, cases are represented in terms of structured attribute-value pairs. Acquisition of cases, either from domain experts or through manually crafting attribute-value pairs from incident reports, constitutes the main reason why CBR systems have not been more common in industries. Manual case generation is a laborious, costlier and time consuming task. Textual CBR (TCBR) is an emerging line that aims to apply CBR techniques on cases represented as textual descriptions. Similarity of cases is based on the similarity between their constituting features. Conventional CBR benefits from employing domain specific knowledge for similarity assessment. Correspondingly, TCBR needs to involve higher-order relationships between features, hence domain specific knowledge. In addition, the term order has also been contended to influence the similarity assessment. This paper presents an account where features and cases are represented using a distributed representation paradigm that captures higher-order relations among features as well as term order information.