Lisansüstü Eğitim Enstitüsü

Permanent URI for this communityhttps://gcris.khas.edu.tr/handle/20.500.12469/77

Browse

Now showing 1 - 20 of 55

Anomaly detection in time series
(Kadir Has Üniversitesi, 2019) Al-Bayati, Taha A.; Ö?renci, Arif Selçuk
The concept of "Internet of Things" is based on connecting any physical object through the internet. This will facilitate our daily lives by dedicating technology in our will. In such a world, the number other interconnected devices is enormous, hence, the need for high performance processing in real-time is huge. This research shines light on the importance of the event processing and machine learning in the time series. A multiple of machine learning algorithms such as support vector machine, decision tree, autoencoder, and K-mean clustering are used for training a time series. A comparison of different methods is analyzed to obtain a robust conclusion about the data. The time series data is used to distinguish the state of emotions for a group of people (15 in total) who participated in an experiment. The state of the emotion may be in one of the four states: stressed, amused, natural, and sad. In this work, we compared the performance of algorithms in terms of their accuracy of predicting the emotions.
Anomaly detection via machine learning
(Kadir Has Üniversitesi, 2023) Çevik, Mesut; Kerestecioglu, Feza; Çevik, Mesut
Retail companies monitor inventory stock levels regularly and manage stock levels based on forecasted sales to sustain their market position. The accuracy of inventory stocks is critical for retail companies to create a correct strategy. Many retail com- panies try to detect and prevent inventory record inaccuracy caused by employee or customer theft, damage or spoilage and wrong shipments. This study is aimed to detect inaccurate stocks using machine learning methods. It uses the real inven- tory stock data of Migros Ticaret A.S¸. of Turkey’s largest supermarket chains. A multiple of machine learning algorithms such as Isolation Forest (IF), Local Outlier Factor (LOF), One-Class Support Vector Machine (OCSVM) were used to detect abnormal stock values. On the other hand, generally, researchers use public data to develop methods, and it is challenging to apply machine learning algorithms to real-life data, especially in unsupervised learning. This thesis shows how to handle real-life data noises, missing values etc. The experimental findings show the perfor- mances of machine learning methods in detecting anomalies in low and high level inventory stock.
Applying machine learning algorithms in sales prediction
(Kadir Has Üniversitesi, 2019) Sekban, Judi; Da?, Hasan
Makine öğrenimi bir çok endüstride üzerinde yoğun çalışmalar yapılan bir konu olmuştur, ve neyse ki şirketler kendi problemlerini çözebilecek çeşitli machine learning yaklaşımları hakkında günden güne daha fazla bilgi sahibi oluyorlar. Fakat, farklı makine öğreniminin modellerinden en iyi şekilde sonuç almak ve verimli sonuçlara ulaşabilmek için, modellerin uygulanış biçimlerini ve verinin doğasını iyi anlamak gerekir. Bu tez, belli bir tahmin görevi için, uygulanan farklı makine öğreniminin algoritmalarını ne kadar iyi sonuç verdiklerini araştırır. Bu amaçla tez, 4 faklı algoritma, bir istifleme topluluğu tekniği ve modeli geliştirmek için belirli bir özelllik seçme yaklaşımı sunar ve uygular. Farklı konfigürasyonlar uygulayarak sonuçlar birbiriyle test edilir. Bütün bu işlemler, gerekli veri önislemeleri ve özellik mühendisliği adımları tamamlandıktan sonra yapılır.
Automatic detection of mgd level (Meibomian gland dysfunction)
(Kadir Has Üniversitesi, 2019) Bütün, Alp Eren; Da?, Tamer
MGD is a common, clinical disease which is the major cause of dry eye. Meibography is a study in order to observe meibomian glands. A meibography image is the infrared photograph of the inner part of eyelids of the patient. Physicians are using meibography images to diagnose MGD. The important result from the infrared meibography image is the level of the disease. Using the level, the appropriate treatment is decided by the physician. Aim of this thesis is to develop the automatic software to detect MGD level from meibography images. Current methods require manual processing takes a long time for detection of MGD level. Thus our aim is to develop an algorithm running in reasonable time and easier to use than the current methods. To achieve our goal, image processing techniques supported by manual editing are used. In addition, some mathematical methods such as drawing line chart, finding local maximum, local minimum, processing image pixels, finding mean, ratio and other methods are used. Manual editing and automatic determination feature is provided to the user of the software. Physician can load a meibography image to the software and can get results in seconds. Result can also be changed. A software using the algorithm described in this thesis is developed and tested. Development of the automatic software to process meibography images for MGD level detection is completed. For testing and valdiation, Unit tests on the code are done. In addition, Bland&Altman method is applied on the results of the software.
Bandwidth allocation and traffic shaping in mobile broadband networks using deep packet inspection
(Kadir Has Üniversitesi, 2015) Arsan, Taner; Arsan, Taner
In this thesis, it is intended to estimate bandwidth and control mobile data usage by utilizing PCC (Policy and Charging Control) function. According to increase in number of mobile devices, data explosion occurs. It is becoming a must to analyze traffic and sharing resources between subscribers according to their usage habits. It is aimed to provide better a better connected world with service assurance by sharing available bandwidth and estimate it to users according to their needs by protocol level and service based QoS. Due to increase in amount of services like Facebook, Twitter, Mobile TV, in general IP networks, providing service assurance becomes more important day by day. That's why the issue of controlling bandwidth is raised. In the most basic sense, system architecture consists of three main components: A cell phone to generate user based traffic, Gateway GPRS Support Node (GGSN) for Deep Packet Inspection (DPI), and Policy and Charging Rule Function (PCRF) for initiating PCC or Non-PCC rules to GGSN according to services that are needed by user. Shortly, the main idea in this thesis is assigning service based QoS to subscribers to provide better service assurance according to their usage. As thought, the reason of preparing this study is to show the dramatical increase in service based traffic, to explain insufficiency in current bandwidth estimation approaches, and the idea of what can be used in the work of providing better service assurance to an end user. PCRF is the best component for providing required bandwidth when they need.
Big data platform development with a telecom DSL
(Kadir Has Üniversitesi, 2013) Bozkuş, Zeki; Bozkuş, Zeki
The amount of data in our world has shown exponential growth in recent years. This creates a very large collection of data sets –so called big data- in many organizations. Enterprises want to process their own big data to generate values from data to improve productivity innovation and customer relationship better than their competitors. However big data is so large and complex that it becomes difficult to process using traditional database management techniques. in this paper we present a system which can be used to analyses for big data of telecom industries. -- Abstract'tan.
Capturing the data similarity among organizations of same nature
(Kadir Has Üniversitesi, 2021) Şenol, Habib; Şenol, Habib
The vertical collaborative clustering aims to unravel the hidden structure of data (similarity) among different sites, which will help data owners to make a smart decision without sharing actual data. For example, various hospitals located in different regions want to investigate the structure of common disease among people of different populations to identify latent causes without sharing actual data with other hospitals. Similarly, a chain of regional educational institutions wants to evaluate their students' performance belonging to different regions based on common latent constructs. The available methods used for finding hidden structures are complicated and biased to perform collaboration in measuring similarity among multiple sites. In this dissertation, the author proposed two approaches of vertical collaborative clustering, namely (1) Vertical Collaborative Clustering Model (2) Vertical Collaborative Clustering based on Bit-Plane Slicing, with superior accuracy over the state of the art approaches. The Vertical Collaborative Clustering Model (V CCM) manages the collaboration among multiple data sites using Self-Organizing Map (SOM). It includes standard procedure and tuning of the exchanged information in specific proportionality to augment the learning process of the clustering via collaboration. Moreover, the VCCM unravels hidden information without compromising the data confidentiality. The aim of the model is to set an ideal environment for the collaboration process among multiple sites. The VCCM is evaluated by purity measurement, using four datasets (Iris, Geyser, Cancer and Waveform). The findings of this study show the significance of the VCCM by comparing the collaborative results with the local results using purity measurement. The VCCM unlocks possible reasons determining impact of collaboration based on related and unrelated patterns. The results demonstrate that the proposed VCCM improves local learning by collaboration and also helps the data owner to make better decisions on the clustering. Additionally, the results obtained have better accuracy than the existing approaches. The proposed Vertical Collaborative Clustering based on Bit-Plane Slicing (VCCBPS) is simple and unique approach with improved accuracy, manages collaboration among various data sites. The VCC-BPS transforms data from input space to code space, capturing maximum similarity locally and collaboratively at a particular bit plane. The findings of this study highlight the significance of those particular bits which fit the model in correctly classifying clusters locally and collaboratively. Thenceforth, the data owner appraises local and collaborative results to reach a better decision. The VCC-BPS is validated by Geyser, Skin and Iris datasets and its results are compared with the composite dataset. It is found that the VCCBPS outperforms existing solutions with improved accuracy in term of purity and Davies-Bouldin index to manage collaboration among different data sites. It also performs data compression by representing a large number of observations with a small number of data symbols. Keywords: Collaborative clustering, Collaboration, Vertical collaborative clustering, Cluster combination, Purity measurement, Similarity measurement
Churn Analysis and Prediction with decision tree and artificial neural network
(Kadir Has Üniversitesi, 2015) Arsan, Taner; Arsan, Taner
Nowadays market competition is increased in logistic sector. Retention of customers due to the competition has gained importance. Retention of customers is more advantageous than gaining new customers. Gaining new customer has 5 units more than retention of existing customer. Especially telecommunication companies use churn analyzing and prediction. However competition increased in different sectors and they need churn analysis. In recent years logistic companies increased in Turkey, so retention of customer and churn analysis are important for logistic companies. Some logistic companies have churn committees and work on customer loyalty. This article includes churn analysis and prediction with decision tree and artificial neural network. In addition, this article includes comparison of 2 different methods for churn analysis. Article results show neural network better than decision for prediction. Because decision tree churn prediction rate is %81, Artificial Neural Networks rate is 97%.
Classification of heart diseases with convolutional neural networks
(Kadir Has Üniversitesi, 2021) Arsan, Taner; Arsan, Taner
Günümüzde kalp hastalıklarının sayısı ve sıklığı artmaktadır. Bu alanda iyileştirmeler yapılabilmesi için yüksek miktarda harcama yapılmaktadır. Kalbin elektriksel iletimindeki atımlar özel cihazlarla kaydedilebilir ve EKG (Elektrokardiyogram) oluşturulabilir. EKG'den üretilen veriler, Taylor Series algoritması ile faz uzaylarına dönüştürülebilir. Kalp hastalığının tespiti için 44 farklı kişiden alınan verilerle MLII sinyallerinden EKG ve faz uzayları oluşturuldu. Bu kayıtların kalp durumunu belirlemek için hem EKG görüntüleri hem de faz uzayı görüntüleri kullanıldı. Kayıtların kalp durumu görüntülere ve sonuçlara Convolutional Neural Networks (CNNs) yöntemi uygulandı ve SVM (Support Vector Machine) algoritması ile karşılaştırılarak başarı oranı ölçüldü. Ayrıca aynı kayıtlar üzerinden eğitim ve test seti değiştirilerek farklı modellerin başarı oranları karşılaştırıldı. EKG ile faz uzayı görüntülerine CNN algoritmasının verdiği sonuçlardaki farklılık tespit edildi. Nowadays, the number and frequency of heart diseases is increasing. High amounts of expenses are incurred in order to make improvements in this area. The beats in the electrical conduction of the heart can be recorded by special devices and ECG (Electrocardiogram) can be created. Data generated from ECG can be transformed into phase spaces with Taylor Series algorithm. In order to determine the detection of heart disease, ECG and phase spaces were created from MLII signals based on 44 different records. Both ECG images and phase space images were used to determine the heart conditions of these recordings. The heart status of the recordings was measured by applying Convolutional Neural Networks (CNNs) method to the images and results compared with the SVM (Support Vector Machine) algorithm. In addition, the success rates of different models were compared by changing the training and test set over the same records. The success rate between ECG and phase space was also determined.
Comparison and analysis of various indoor positioning systems techniques
(Kadir Has Üniversitesi, 2016) Demirkol, Derya; Da?, Tamer
indoor Positioning has been a research subject in order to facilitate people life easier. Different type of methods has been implemented and tested by the years. The Global Positioning System (GPS) is a satellite based navigation system. This technique is using outdoor environment to navigate people or buildings. in indoor positioning GPS signals are usually too weak to provide accurate positioning estimate. Other technique need to investigate to get better result of accuracy. Ultrasonic Positioning Systems RFiD Computer Vision System RF Wireless indoor Positioning System has been used. Recent year Wireless indoor Positioning Technique is most popular technique. it is easier to set up every indoor environment by using Access Points (AP) and costs are very low comparing to other techniques. Also Wireless technique does not need any extra component or effort. in this thesis indoor Positioning techniques were investigated. Different algorithms were implemented to estimate location in indoor areas and accuracy comparison has been made by using Wireless technology.
Concept map based learning system
(Kadir Has Üniversitesi, 2014) Çolak, Duygu; Ö?renci, Arif Selçuk
Concept Map is a training technique which accelerates facilitates and consolidates the learning process. Concept map uses concepts and connections between concepts to train and transmit the information during education. A web application which is based on this technique has been developed. The application is designed for the teachers from different branches students for various ages and everyone who wants to teach and learn any subject. The application is web based and designed for different screen sizes so users may access the application via their mobile devices for instance mobile phones or tablets and personal computers. Tutor adds course content determines the concepts and relations from the content and connects them on drawing area in order to form a map. Student chooses one of the courses reads the course content and tries to draw the concept map by deciding on concepts and relations. The application gives feedback to the student while drawing the map which has been defined by the course teacher thus facilitates the drawing and learning. Student may compare his concept map with the teachers’ thus continues the drawing operation until the learning process completes. -- Abstract'tan.
Design and implementation of a mobile prescription system with patient-healthcare professional interaction
(Kadir Has Üniversitesi, 2015) Dağ, Tamer; Dağ, Tamer
In this thesis, a mobile prescription reminder and scheduling system application with patient-healthcare professional interaction is designed and implemented. By using the application, different types of users are able to create and manage prescriptions and are reminded to take the medication based on the prescription data. Besides, users are able to reach the prospectuses of the drugs. One of the most important functionality of the application is to provide patient-healthcare professional interaction. By using the application, healthcare professionals can assign prescriptions to their patients and they can also monitor their medicine compliance. Since the medicine compliance is very important to have an effective treatment, the main goal of this application is to help people to take their medicines on time and find the prospectus information of the drugs easily by using their mobile phones.
Developing novel techniques for spatial domain LSB i̇mage steganography
(Kadir Has Üniversitesi, 2019) Shehzad, Danish; Da?, Tamer
Steganography is one of the most noteworthy information hiding mechanism, which is used as an alternative to cryptography in order to provide adequate data security. Image steganography is one of the key types of steganography where a message to be transmitted is hidden inside a cover image. The most commonly used techniques for image steganography rely on LSB Steganography. In this thesis, new techniques are developed for LSB image steganography to achieve maximum security, optimal data capacity along with provisioning of efficient steganography mechanism. In the first part of this work, a novel technique based on pairs matching is developed for LSB image steganography. In this technique MSBs along with LSBs are used in a delicate method for data hiding for the first time. The message bits from the secret information are compared with all defined pixel pairs and replace the least two significant bits with respective matched pair number. This technique shows good quality of stego image along with adequate peak signal to noise ratio and provides high payload of secret message. In the second part, threshold-based LSB image steganography technique is developed. This technique also works in spatial domain and categorizes the pixels based on threshold defined categories. Maximum four bits and minimum one bit is embedded in pixel based on its category. The prominence in THBS is on security and payload as it uses bits proficiently for data embedding. ETHBS allows efficient execution of the algorithm along with provisioning of optimal security. In the last part 1LSB Image steganography technique based on blocks matrix determinant is developed. It is a technique in which data is embedded by making minimal changes in image pixels. This technique is 1LSB substitution technique that works on matrix determinant of 2 by 2 blocks of image pixels. This technique ensures high PSNR and ensures good quality of stego image.
Development of hybrid MPI+UPC parallel programming model
(Kadir Has Üniversitesi, 2011) Bozkuş, Zeki; Bozkuş, Zeki
Parallel Computing is a form of computation that divides a large set of calculations into tasks and runs on multi-core machines simultaneously. Today, Message Passing Interface (MPI) is the most widely used parallel programming paradigm that provides programming both for symmetric multi-processors (SMPs) which consists of shared memory nodes with several multi-core CPUs connected to a high speed network and among nodes simultaneously. Unified Parallel C (UPC) is an alternative language that supports Partitioned Global Address Space (PGAS) that allows shared memory like programming on distributed memory systems.In this thesis, we describe the MPI, UPC and hybrid parallel programming paradigm which is designed to combine MPI and UPC programming models. The aim of the hybrid model is to utilize the advantages of MPI and UPC; these are, MPI?s data locality control and scalability strengths with UPC?s global address space, fine grain parallelism and ease of programming to achieve multiple level parallelism. This thesis presents a detailed description of hybrid model implementation comparing with pure MPI and pure UPC implementations. Experiments showed that the hybrid MPI+UPC model can significantly provide performance increases up to double with pure UPC implementation and up to 20% increases in comparison to pure MPI implementation. Furthermore, an optimization was achieved which improved the hybrid performance an additional 20%.
Dynamic load balancing in distributed systems: “hands of god” in parallel programming with MPi
(Kadir Has Üniversitesi, 2014) Arsan, Taner; Arsan, Taner
in today’s technology what really is missing in computer systems is more artificial intelligence and in the same time implanting lots of intelligence in computer systems is not as easy as it sounds but even one step ahead to make computer software to act more efficient and intelligent is noteworthy. MPiCH is a message passing interface framework designed to be the host for parallel programs but like too many other great programming frameworks development of MPiCH is ongoing and once a while we are witnessing new updates which mostly these updates are in order to support more functionality and performance improvement updates are more rare. One of the issues that happened while we were working on Snake in the box problem (more details on this problem could be found in Wikipedia) with my parallel programming Professor Dr Turgay Altilar in Kadir Has University was lack of intelligence in the parallel version of the algorithm of Snake in the box in MPiCH framework. -- Abstract'tan.
Dynamic multi threshold priority packet scheduling algorithms for wireless sensor networks
(Kadir Has Üniversitesi, 2015) Dağ, Tamer; Dağ, Tamer
Kablosuz sensör ağlarında farklı türlerde özellikle gerçek zamanlı ve gerçek olmayan zamanlı paket zamanlama gereklidir. Sensörlerin enerji kullanımlarını ve iletim gecikmelerini azaltmak önemlidir. Tezimde yeni paket zamanlama algoritmalarını geliştirerek bunu kablosuz sensör ağlarına entegre etmeye çalışarak enerji kullanımını ve iletim gecikmelerini geliştirerek daha verimli yapıyorum. Tasarladığım dinamik çoklu eşik ve öncelikli paket zamanlama algoritmaları, düşük öncelikli veriler için gecikme zamanını ve veri kaybını azaltarak bunu yüksek öncelikli verilere adil bir şekilde davranarak yapıyor. Eşik algoritmaları günümüzde en çok kullanılan paket zamanlama algoritmalarıyla kıyaslanıyor. Bunlar ilk gelen ilk servis edilir algoritması ile öncelikli paket zamanlama algoritmasıdır. Simülasyon sonuçları gösteriyor ki dinamik çoklu eşik ve öncelikli paket zamanlama algoritmaları düşük öncelikli verilerin servis kalitesini arttırıyor ve bunu yüksek öncelikli verilerin servis kalitesini koruyarak yapıyor.
Energy trading in Turkish power market through options bileteral agreements and spot market
(Kadir Has Üniversitesi, 2013) Tasdemir, Ecem Ezgi; Aya?, Zeki
in today’s world electricity energy which has a vital place in energy market is irreplaceable in any field because of being the basic entry in production factors and because of continuous increasing demand to electricity thanks to technological development. Especially in our country which has a rapid acceleration for the last decade by reorganizing Electricity Energy Market it is aimed to build a competitive structure in production and retail transactions. Also with the process of development in Turkey the structure of the market changed varied and created different alternatives for efficient buying-selling transactions. in this work we built a model on the purpose to find the solution of the most efficient scenarios and maximum efficiency by using electricity spot market which has transactions with a determined capacity limit derivatives market and bilateral agreements. The amount of efficiency is calculated by using the Model for each 100 different power scenarios’ in Excel and the results are got and interpreted. As a result of this the most efficient scenario determined. -- Abstract'tan
Feature selection and discretization for improving classification performance on CAC data set
(Kadir Has Üniversitesi, 2013) Dağ, Hasan; Dağ, Hasan
Data Mining usage in Health Sector increased much in this decade because of the need for efficient treatment. From cost-cutting in medical expenses to acting as a Decision Support System for patient diagnosis, Data Mining nowadays is a strong companion in Health Sector. The dataset used in this thesis belongs to Dr. Nurhan Seyahi. Dr. Nurhan Seyahi and his colleagues made a research about Coronary Artery Calcification in 178 patients having renal transplantation recently. They used conventional statistical methods in their research. By using the power of data mining, this thesis shows the importance of feature selection and discretization used with classification methods for acting as a decision support system in patient diagnosis for CAC Dataset. Just by looking at seven important attributes, which are; age, time of transplantation, diabetes mellitus, phosphor, rose angina test, donor type and patient history, doctors can decide whether the patient has coronary artery calcification or not with approximately 70% accuracy. After the discretization process this accuracy approximately increases to 75% in some algorithms. Thus becoming a strong decision support system for doctors working in this area.
Global alignment of metabolic pathways and protein-protein interaction networks
(Kadir Has Üniversitesi, 2014) Erten, Cesim; Erten, Cesim
Metabolic pathways and protein interaction networks are essential at almost every function for living organisms. Simply, while reactions produce life energy within cells, protein interaction networks provide biological functions. Also, abnormal reactions or interactions cause various disorders. Thus, in bioinformatics, most of the studies are based on these networks in order to find hopeful results for these disorders and biological challenges. Solving alignment problem is one of these studies such that it tries to find similar reactions, proteins or functions. In this thesis, we focus on that problem within both metabolic pathways and protein interaction networks. Firstly, we propose a constrained alignment algorithm, CAMPways, for one-to-many alignment of metabolic pathways and we extend the framework, CAPPI, for one-to-one protein interaction network alignment with necessary changes. Afterwards, we provide the computational intractability of the problem and finally we compare our algorithm with different algorithms on actual metabolic pathways and protein interaction networks.
Global many- to - many aligment of multiple protein-protein interaction networks
(Kadir Has Üniversitesi, 2013) Alkan, Ferhat; Aşıcı, Tınaz Ekim
Proteins are essential parts of organisms and almost every biological process within a living cell is mediated by proteins and their interactions. Due to such importance, proteins are at the core of many researches in systems biology and evolutionary biology. In particular, defining the function of a protein and identfying functionally orthologous proteins are crucially important in many research areas and precise function of a protein can only be defined by biochemical and structural studies. However, many computational methods are also developed for such purposes and they use the sequence and interaction data of proteins since it provides a presumption about the chemical structure of a protein. For example, network alignment studies aims to find clusters of functionally related proteins across given protein interaction networks usually by implementing the given networks as graphs and employing some graph theoretical approaches. In this thesis, we focused on the problem of global many-to-many alignment of multiple protein-protein interaction networks. We define the problem as an optimization problem and this is the first combinatorial definition that is given for the problem in the literature. Then, we prove the computational intractability of this problem and we propose a new heurictic algorithm for the solution. We test the proposed algorithm BEAMS on both actual and synthetic PPI networks and it outperforms the existing algorithms, that serve at similar purpose, in terms of many evaluation aspects.

Browse

Browsing Lisansüstü Eğitim Enstitüsü by Department "Enstitüler, Lisansüstü Eğitim Enstitüsü, Bilgisayar Mühendisliği Ana Bilim Dalı"