Majlesi Journal of Multimedia Processing
Volume:5 Issue: 2, Jun 2016

Fundamental changes have occurred in social interactions of the people with the advent and expansion of online social networks. With the expansion of social networks and the ever-increasing number of their users, the prediction of the users' relationships has turned into a difficult and complicated problem in these networks. Link prediction examines the links missing on the current network as well as the new links created in the future in social networks. Supervised and unsupervised methods can be used to predict the link. In unsupervised link-prediction method, the ranking of pair nodes is done only using one criterion, and in contrast to supervised link-prediction methods, they can complete the information from multiple scales and usually make real-world network model better. One of the methods proposed recently by Wang et al. states the problem of link prediction within the framework of supervised link prediction. This framework includes a re-weighing scheme based on the extracted features from high-profile interactions patterns across the network with great performance in link prediction, but in some supervised methods, it performs poorly not improving the precision of the link prediction. Thus, to solve the problem of link prediction, we introduce a new supervised link-prediction framework. Using the graph-edge clustering, supervised learning, and feature selection with EAs such as genetic algorithm (GA) in heterogeneous social networks, the link prediction problem was used. Furthermore, AdaBoost algorithm was applied to train learning models. In doing so, the DBLP scientific dataset was used. The results showed that feature selection using evolutionary GA improves the link prediction precision in social networks related to DBLP scientific publications by 5% in the logical regression model and 100% in the neural network and naive Bayes models. However, in the randomized forest model, precision is reduced by about 20%.

Social networks consist of much knowledge about the associations between human beings. However, unhappily, the kind of associations between human beings is concealed in these networks and majority of associations stay without title. Some social network sites clarify an option for defining this association which should be completed by user in order to find a solution for this problem but this leads network enters specific associations which are clear and unclear. Thus, a new challenge named forecasting of the connection kind is presented that intends to recognize the kind of association among entities. At first, the significance of forecasting of connection kind in social networks and their benefits have been debated in this paper. Then, an enhanced algorithm is offered so that it can raise the preciseness of resolving this issue that is rules-based and time-limited and it leads to decrease the number of potential goal edges by filtrating the non-goal edged and then it distinguishes the goal edges by fetch of edges with the most chance for each node and causes to raise preciseness of forecasting the link kind. Finally, in order that it appraises the indicated method, a data set of scientific essays has been applied for researching the kind of the relationships between authors that this data set is the subset of DBLP network authors. The outcomes of the estimation of this procedure indicate raised preciseness and exactness in comparison with the base.

In hyperspectral images, each pixel appears as a set of mixed spectral vectors. Estimating the number of reference substances, also called endmembers, their spectral signatures, and their abundance fractions are main challenges encountered in the hyperspectral image processing. This paper introduces a new unsupervised algorithm, called Rotational algorithm for this purpose. Rotational algorithm is a geometric method that assumes existence of at least one pure pixel per each endmember in the image. These pure pixels would be vertices of a simplex. Rotational algorithm makes use of these facts: in a simplex, each vector having at least one extreme component is definitely a vertex; Also the affine transformation of a simplex is a simplex. To evaluate the performance of the Rotational algorithm, we compared its results with the popular VCA algorithm. By comparing experimental results, it was found that although the running time of the Rotational algorithm is a little more than the VCA algorithm, but its accuracy in determining the endmembers has improved in comparison to the VCA algorithm.

Due to the rapid and increasing progress of medical equipment and medical imaging machines, a large number of digital images are produced in therapeutic centers and stored in the large databases. Retrieving a set of images which are most similar to a query image is a major challenge in this field. A popular way to address this problems is to apply image retrieval based on a bag of visual words. A popular technique for producing visual words is to use K-means clustering algorithm. However, the effectiveness of K-means algorithm highly depends on the initial selection of cluster centroids which are selected randomly in the original K-means algorithm. Therefore, the image retrieval task is affected by poor clustering solutions due to random selection of cluster centroids. The goal of this paper is to overcome this problem and improving the accuracy of content based image retrieval from medical image databases using the optimal selection of initial points in K-means clustering algorithm. In the proposed method, after extracting SIFT features from the color images, the optimal centroid points are selected based on different ranges, weights and means of samples and visual words are created based on the selected centroid. The result of experiments on applying three different algorithms for selecting initial cluster centroids show that selecting the optimized initial points results in producing more discriminative visual words and provide favorable accuracy for medical image retrieval systems.

Classifying audio signals to groups such as music and speech is significant for many multimedia document recovery systems. So far, several approaches have been proposed for categorizing audio data as music or speech. In this paper, two approaches have been taken for discriminating and classifying these data; in one of them audio signals have a binary classification of speech/music and in another there is a ternary classification of speech/music/mixture. After extracting some audio features (such as low short-time energy ratio of frames, standard deviation of the zero-crossing rate, mean and variance of the spectral flux, mean and variance of the discrete wavelet transform, mean of the linear predictive coefficients, mean of the Mel frequency cepstral coefficients), the MLP neural network was used to classify them. Simulations on the network showed a classification accuracy of 98.7% for the binary classification and 93.3% for the ternary one. This confirms the usefulness of the proposed method.