Sasan Karamizadeh

Advantage and drawback of support vector machine functionality

Support Vector Machine(SVM)is one of the most efficient machine learning algorithms, which is mostly used for pattern recognition since its introduction in 1990s. SVMs vast variety of usage, such as face and speech recognition, face detection and image recognition has turned it into a very useful algorithm. This has also been applied to many pattern classification problems such as image recognition, speech recognition, text categorization, face detection, and faulty card detection.Statistics was collected from journals and electronic sources published in the period of 2000 to 2013. Pattern recognition aims to classify data based on either a priori knowledge or statistical information extracted from raw data, which is a powerful tool in data separation in many disciplines. The Support Vector Machine (SVM) is a kind of algorithms in biometrics. It is a statistics technical and used orthogonal transformation to convert a set of observations of possibly correlated variables into a set of values of linearly uncorrelated variables.

Identifying Benefits and Risks Associated with Utilizing Cloud Computing

Cloud computing is an emerging computing model where IT and computing operations are delivered as services in highly scalable and cost effective manner. Recently, embarking this new model in business has become popular. Companies in diverse sectors intend to leverage cloud computing architecture, platforms and applications in order to gain higher competitive advantages. Likewise other models, cloud computing brought advantages to attract business but meanwhile fostering cloud has led to some risks, which can cause major impacts if business does not plan for mitigation. This paper surveys the advantages of cloud computing and in contrast the risks associated using them. Finally we conclude that a well-defined risk management program that focused on cloud computing is an essential part of gaining value from benefits of cloud computing.

PATTERN RECOGNITION TECHNIQUES: STUDIES ON APPROPRIATE CLASSIFICATIONS

Pattern recognition techniques are divided into categories of supervised, unsupervised and semi supervised. Supervised pattern recognition methods are utilized in the examination of various sources’ chemical data such as sensor measurements, spectroscopy, and chromatography. The unsupervised classification techniques use algorithms to classify and analyze huge amounts of raster cells. Semi-Supervised Learning is an approach that is in the middle ground between supervised and unsupervised learning and guarantees to be better at classification by involving data that is unlabeled. In this paper, we tried to categories pattern recognition methods and explain about each of them and we compared supervised method with unsupervised method in terms of types and location of features.

Statistical and data mining methods in credit scoring

The growing interest in the credit industry resulted in credit scoring being developed as an essential component, especially in the credit department of banks that deals with huge sums of credit data. When a bank or a credit corporation is assessing a credit application request, they will have to decide whether to approve or deny it. This necessitates the utilization of credit scoring. Although pioneers attempt to compensate for risks via interest rates, current investigations on financial conditions of different sections of society confirmed that interest could not replace risk assessment, which means that credit risk requires its own specialized assessment. With the assistance of sorting methods, credit scoring simplifies the decision-making process. It is almost impossible to analyze this large amount of data in the context of manpower and economy, although the data mining technique helps alleviate this complexity. Nowadays, there are a lot of data mining methodologies being utilized in the management of credit scoring. However, each method has its advantages and limitations, and there has not been a comprehensive approach in determining the most utilized data mining technique in the context of credit scoring. The major goal of this paper is to provide a complete literature survey on applied data mining methods, such as discriminant analysis, logistic regression, K-nearest neighbor, Bayesian classifier, decision tree, neural network, survival analysis, fuzzy rule-based system, support vector machine, and hybrid methods. These findings will assist researchers in realizing the most suitable approach in evaluating credit scores, pinpoint limitations, enhance them, and propose new approaches with improved capabilities. Finally, the limitations of the new approaches are discussed, and further suitable methods are recommended.

Improve TCP performance over mobile ad hoc network by retransmission timeout adjustment

Conventional TCP suffers from inability to detect frequent link failure in MANET and consequently could not adjust RTO for reconstructed route. Packet losses due to link breakage must be differentiated from congestion loss to discover link breakage. Moreover, after link breakage, retransmission timeout in standard TCP becomes too long due to successive back-off executions. Using this long RTO for rebuilt route forces sender to remain idle unnecessarily in case of future packet loss. In this paper, a new End-to-End approach which entirely works in transport layer is proposed to improve TCP performance. It distinguishes link failure loss from congestion loss based on fluctuation in history of queue usage rate. Ascending growth of queue usage value intensifies probabilities of congestion losses while averaged queue usage values around fixed value can be sign of link failure losses. After temporary link failure detected and packets from rebuilt route received, new approach compares characteristics of reestablished route with broken route in term of Relative On-way Trip Time and number of Hops. Then sender decides to increases RTO to prevent packet retransmission overhead or decreases RTO to resume transmission during unnecessary idle time. Simulation results which are done under different conditions by ns2 illustrated that new approach enhances TCP performance up to 10%.

Face Recognition via Taxonomy of Illumination Normalization

Presently, the difficulty in managing illumination over the face recognition techniques and smooth filters has emerged as one of the biggest challenges. This is due to differences between face images created by illuminations which are always bigger than the inter-person that usually be used for identities’ recognition. No doubt, the use of illumination technique for face recognition is much more popular with a greater number of users in various applications in these days. It is able to make applications that come with face recognition as a non-intrusive biometric feature becoming executable and utilizable. There are tremendous efforts put in developing the illumination and face recognition by which numerous methods had already been introduced. However, further considerations are required such as the deficiencies in comprehending the sub-spaces in illuminations pictures, intractability in face modelling as well as the tedious mechanisms of face surface reflections as far as face recognition and illumination concerned. In this study, few illuminations have been analyzed in order to construct the taxonomy. This covers the background and previous studies in illumination techniques as well the image-based face recognition over illumination. Data was obtained from the year of 1996 through 2014 out of books, journals as well as electronic sources that would share more on the advantageous and disadvantageous, the current technique’s performance as well as future plan.

Packet loss differentiation of TCP over mobile ad hoc network using queue usage estimation

The main deficiency of standard TCP and all existing variations in ad hoc mobile network arises in its inability to classify packet losses. Conventional TCP diagnoses all losses due to congestion since non-congestion losses are rare and ignorable in wired network. However, these losses are generated frequently in MANET due to signal attenuation, interference and nodes mobility. In order to prevent TCP form invoking congestion control over non-congestion loss and vice versa, Packet loss classification with high accuracy is desirable. This paper proposes an end-to-end sender-side approach which aims to classify congestion loss, wireless channel loss and link failure loss using queue usage estimation. Relative One-way Trip Time is key factor in calculating queue usage. Packet losses with queue usage rate greater than predefined threshold is always interpreted as congestion loss no matter how loss recognized. However, queue usage less than threshold signify presence of non-congestion losses. Non-congestion loss which recognized by three duplicate ack labels loss due to wireless channel error since it is indication of route existence between communicating end point that duplicate ACKs moved along. However Non-congestion loss which recognized after retransmission timer expiration has been probably due to link failure. Different approaches efficiency are evaluated based on loss classifications accuracy under different conditions such as various flow numbers, various wireless loss rates and variable speed for nodes. Simulations illustrate that obtained accuracy of enhanced scheme is the highest among proposed methods.

The application of fusion of heterogeneous meta classifiers to enhance protein fold prediction accuracy

Protein fold prediction problem is considered as one of the most challenging tasks for molecular biology and one of the biggest unsolved problems for science. Recently, varieties of classification approaches have been proposed to solve this problem. In this study, a fusion of heterogeneous Meta classifiers namely: LogitBoost, Random Forest, and Rotation Forest is proposed to solve this problem. The proposed approach aims at enhancing the protein fold prediction accuracy by enforcing diversity among its individual members by employing divers and accurate base classifiers. Employed classifiers combined using five different algebraic combiners (combinational policies) namely: Majority voting, Maximum of Probability, Minimum of Probability, Product of Probability, and Average of probability. Our experimental results show that our proposed approach enhances the protein fold prediction accuracy using Ding and Dubchaks dataset and Dubchak et al.s feature set better than the previous works found in the literature.

An overview of objectionable image detection

Rapid growth of internet made sharing and distribution of media files became easier. This phenomena and this fact that internet is boundary less network, raised the concern of exposure to unwanted images and more importantly offensive images. Pornographic images are among most offensive images especially for minors. Although there are numbers of researches conducted to address this challenge, but this field is still crude. In this paper, different approaches to pornographic image detection is reviewed and categorized by focusing to content based methods. Reviewing these papers, general limitations of these works are discussed including lack of standard dataset for this field and standard definition which is well accepted for pornographic or naked images.