Vol. 31 · No. 6+

The paper describes the possibilities of using Fourier-Mellin transform for registering images of holographic interferograms. Registered holographic images will then allow automating their evaluation. Registration based on changes in image intensities using the discrete integral transforms was selected from the methods of registration. Whereas it was necessary to register the images, which are not only translated, but also rotated and with the change of scale, the Fourier-Mellin transform was used. Use of the image discrete transforms is original in this field, the proposed processing algorithm also contains simplified mean of calculating the angle of rotation of the test image instead of common Fourier-Mellin transformation method sequence.

As the original rough set model is quite sensitive to noisy data, Ziarko proposed the variable precision rough set (VPRS) model to deal with noisy data and uncertain information. This model allowed for some degree of uncertainty and misclassification in the mining process. In this paper, the variable precision rough set model for an incomplete information system is proposed by combining the VPRS model and incomplete information system, and the beta-lower and beta-upper approximations are defined. Considering that classical VPRS model lacks a feasible method to determine the precision parameter beta when calculating the beta-reducts, we present an approach to determine the parameter beta. Then, by calculating discernibility matrix and discernibility functions based on beta-lower approximation, the beta-reducts and the generalized decision rules are obtained. Finally, a concrete example is given to explain the validity and practicability of beta-reducts which is proposed in this paper.

In BDD-based fault tree analysis, the size of BDD encoding fault trees heavily depends on the chosen ordering. From a theoretical point of view, finding the best ordering is an intractable task. So, heuristics are used to get good orderings. The most simple, and often one of the best heuristics is depth first left most (DFLM) heuristic. Although having been used widely, the performance of DFLM heuristic is still only vaguely understood, and not much formal work has been done. This paper starts from two different research objects: fault tree without repeated events (NRFT) and fault tree with repeated events (RFT). For NRFT, the BDD generated according to DFLM ordering is proved to be the smallest BDD with the size equal to the total number of events. For RFT, a randomized algorithm is firstly proposed to create reliable benchmarks including large number of random fault trees with different specificities. Then, these benchmarks are used to perform two types of experiments to study the performance of DFLM heuristic. For RFT with small number of repeated events, it is found that the sizes of the BDD built over DFLM orderings are only slightly larger than the sizes of the RFT with different specificities. However, with the increase of the number of repeated events, we encounter the size explosion problem, and the change of repeated event distribution patterns will have a significant impact on the sizes of the BDD built over DFLM orderings. We also find that the number of repeated events is the more important measure than some other specificities (shape, logical type of top gate and OR/AND gate distribution) to estimate the level of the difficulty in BDD-based fault tree analysis.

Multidimensional Scaling (MDS) is an important class of techniques for embedding sets of patterns in Euclidean space. Most often it is used to visualize in mathbbR3 multidimensional data sets or data sets given by dissimilarity measures that are not distance metrics. Unfortunately, embedding n patterns with MDS involves processing O(n2) pairwise pattern dissimilarities, making MDS computationally demanding for large data sets. Especially in Least Squares MDS (LS-MDS) methods, that proceed by finding a minimum of a multimodal stress function, computational cost is a limiting factor. Several works therefore explored approximate MDS techniques that are less computationally expensive. These approximate methods were evaluated in terms of correlation between Euclidean distances in the embedding and the pattern dissimilarities or value of the stress function. We employ Procrustes Analysis to directly quantify differences between embeddings constructed with an approximate LS-MDS method and embeddings constructed with exact LS-MDS. We then compare our findings to the results of classical analysis, i.e. that based on stress value and correlation between Euclidean distances and pattern dissimilarities. Our results demonstrate that small changes in stress value or correlation coefficient can translate to large differences between embeddings. The differences can be attributed not only to the inevitable variability resulting from the multimodality of the stress function but also to the approximation errors. These results show that approximation may have larger impact on MDS than what was thus far revealed by analyzes of stress value and correlation between Euclidean distances and pattern dissimilarities.

This paper describes our contribution to the research of parametrized models and methods for detection and prediction of significant meteorological phenomena, especially fog and low cloud cover. The project covered methods for integration of distributed meteorological data necessary for running the prediction models, training models and then mining the data in order to be able to efficiently and quickly predict even sparsely occurring phenomena. The detection and prediction methods are based on knowledge discovery – data mining of meteorological data using neural networks and decision trees. The mined data were mainly METAR aerodrome messages, meteorological data from specialized stations and cloud data from special airport sensors – laser ceilometers.

In this paper, a novel algorithm for feature extraction, named supervised kernel locally principle component analysis (SKLPCA), is proposed. The SKLPCA is a non-linear and supervised subspace learning method, which maps the data into a potentially much higher dimension feature space by kernel trick and preserves the geometric structure of data according to prior class-label information. SKLPCA can discover the nonlinear structure of face images and enhance local within-class relations. Experimental results on ORL, Yale, CAS-PEAL and CMU PIE databases demonstrate that SKLPCA outperforms EigenFaces, LPCA and KPCA.

In the application layer multicast (ALM), clustering nearby nodes can effectively improve the multicast performance. However, it is difficult for the ALM solution to quickly and accurately position the newcomer, because group members have no direct knowledge of underlying network topology. Additionally, ALM delivery trees with different performances are built when group members join the group in different join sequences. To alleviate the above problems, this paper proposes a distance-heuristic tree building protocol (called DHTB). DHTB uses our proposed distance-constrained cluster model and close-member-first-receive (CF) rule. In the model, most nearby nodes are grouped into some distance-constrained clusters, with little cluster organization and maintenance overhead. The CF rule arranges or rearranges the locations of group members according to related distances, and effectively positions the newcomer with the help of on-demand landmarks. Both the distance-constrained cluster model and CF rule are distance-heuristic. Therefore DHTB can alleviate the join sequence problem, and build the ALM tree with desirable performance.

Most learning algorithms are designed to work on a single dataset. However, with the growth of networks, data is increasingly distributed over many databases in many different geographical sites. These databases cannot be moved to other network sites due to security, size, privacy, or data ownership consideration. In this paper, we propose two decomposable versions of Naive Bayes Classifier for horizontally and vertically partitioned data. The goal of our algorithms is to achieve the learning objectives for any data distribution encountered across the network by exchanging minimum local summaries among the participating sites.

Nowadays, almost all text corpora, such as blogs, emails and RSS feeds, are a collection of text streams. The traditional vector space model (VSM), or bag-of-words representation, cannot capture the temporal aspect of these text streams. So far, only a few bursty features have been proposed to create text representations with temporal modeling for the text streams. We propose bursty feature representations that perform better than VSM on various text mining tasks, such as document retrieval, topic modeling and text categorization. For text clustering, we propose a novel framework to generate bursty distance measure. We evaluated it on UPGMA, Star and K-Medoids clustering algorithms. The bursty distance measure did not only perform equally well on various text collections, but it was also able to cluster the news articles related to specific events much better than other models.

Most of the existing semi-supervised clustering algorithms depend on pairwise constraints, and they usually use lots of priori knowledge to improve their accuracies. In this paper, we use another semi-supervised method called label propagation to help detect clusters. We propose two new semi-supervised algorithms named K-SSMST and M-SSMST. Both of them aim to discover clusters of diverse density and arbitrary shape. Based on Minimum Spanning Tree's algorithm variant, K-SSMST can automatically find natural clusters in a dataset by using K labeled data objects where K is the number of clusters. M-SSMST can detect new clusters with insufficient semi-supervised information. Our algorithms have been tested on various artificial and UCI datasets. The results demonstrate that the algorithm's accuracy is better than other supervised and semi-supervised approaches.

Web services are an emerging paradigm which aims at implementing software components in the Web. They are based on syntactic standards, notably WSDL. Semantic annotation of Web services provides better qualitative and scalable solutions to the areas of service interoperation, service discovery, service composition and process orchestration. Manual annotation is a time-consuming process which requires deep domain knowledge and consistency of interpretation within annotation teams. Therefore, we propose an approach for semi-automatically annotating WSDL Web services descriptions. This is allowed by Semantic Web Service Engineering. The annotation approach consists of two main processes: categorization and matching. Categorization process consists in classifying WSDL service description to its corresponding domain. Matching process consists in mapping WSDL entities to pre-existing domain ontology. Both categorization and matching rely on ontology matching techniques. A tool has been developed and some experiments have been carried out to evaluate the proposed approach.

Traditional approaches in data mining focus on support and confidence measures which are just statistics based. Support and confidence measures which are based on the frequency count of the items enable us to derive the frequent itemsets. The frequency of the items as a single factor does not represent the interestingness of the items. To enhance the process of data mining tasks based on the value of the product, several researches were conducted. It resulted in utility mining which is an emerging field of research in data mining. In the recent years various data mining approaches have been implemented in order to find the high utility itemsets. The main objective of utility mining is to identify the itemsets with highest utilities, by considering the subjectively defined utility values, as set by the user. Existing methods based on utility mining concept focus on centralized systems where the data and associated processing is pertained to a particular location. As a further step ahead we try to implement the utility mining concept in a distributed environment. In this approach we use a sophisticated way of mining high utility itemsets using a Fast Utility Mining (FUM) algorithm.