2,462
Archived Theses
0
DOIs Assigned
0%
DOI Rate
Discipline
Using text representation and deep learning methods for turkish text classification
The heavy use of the Internet has led to a significant increase in the amount of text content produced in online platforms. Huge amount of online textual data is difficult to process, and new techniques have begun to be developed to process online data automatically. New word and document representation methods and deep learning-based classifiers have emerged recently to work with large text datasets as an alternative way to traditional text processing methods. The vast majority of studies using these methods were done with English texts. For Turkish texts, these methods have been used in the last 2 or 3 years. In this thesis, our aim is to evaluate the performances of new text representation and deep learning-based methods on classification of Turkish texts having different characteristics to show the usability of these methods on different document types. Therefore, these methods are used for the problems of sentiment and document classification and their performances are compared with traditional text classification methods. In order to make performance comparisons of the classifiers for the two text classification tasks that studied, deep learning-based convolutional neural networks and long short-term memory networks are used; as well as traditional classifiers which frequently used for Turkish texts in the literature. In the experimental evaluations it is found that embedding methods have similar performance with the traditional tf and tf-idf weighting methods, and in some cases achieve higher classification success. Deep learning-based classifiers have equal or higher classification success than the traditional classifiers.
A hybrid approach for feature reduction
In this thesis, currently available feature selection and dimension reduction methods are analyzed on three different datasets. Particular emphasis is given to filtering-based feature selection methods, which are popular in machine learning area due to their speed. In addition, a linear generalization of the features is obtained by applying dimension reduction in case the scores of the features are close to each other. The feature selection methods that are examined in this thesis evaluate features individually by ignoring correlation between them. At this stage, features marked as redundant are accepted completely irrelevant. Hence, in this thesis, a hybrid feature reduction approach is proposed. In this approach, unchosen features have also opportunity to be involved in classification and clustering by providing a linear projection. Hence, with this approach both the most relevant features selected by feature selection algorithm and projection of unchosen features are utilized. The results obtained in terms of f-measurement show that, besides selecting the best feature subset consisting of top-n features, giving a perfect-fit chance to unchosen features may lead to satisfactory classification results.
Development of a virtual reality exposure therapy system: An example of obsessive- compulsive disorder
With the developing technology, virtual reality technology has started to take its place in our lives. This technology, which has found its place in many fields from health to law, education to tourism, offers innovative solutions in different sectors. Virtual reality glasses have started to be used in the field of psychotherapy, keeping pace with the advancing technology. Virtual Reality Exposure Therapy (VRET) offers a more accessible, convenient and safer environment than traditional exposure therapy in the treatment of many anxiety disorders such as phobias. While many VRET solutions tout their efficacy in therapy, there has been little large-scale uptake. There are no guidelines for developing these solutions, so there is no standardization across solutions. There is therefore a need for safe, commercially viable VRET solutions that are practically effective and can be used in most exposure therapies. Based on this, we have developed a scalable and manageable VRET application. This study offers an innovative solution from both a clinical and technological point of view, presenting a generic VRET framework that is not only limited to a specific domain, such as obsessive-compulsive disorder, but is also applicable to other anxiety disorders. Furthermore, a system has been developed that allows the expert to actively follow the process and customize the therapy according to individual needs. In these aspects, our study is considered as an important step that can promote the widespread use of VRET.
An educational game design for robotic coding
As technological advancements continue to evolve, access to new learning methods has become increasingly diverse. Learning is no longer confined to books or newspapers but extends to various digital platforms. These platforms aim to make learning more accessible, engaging, and memorable. Instead of traditional, monotonous learning methods, integrating educational content with interactive and entertaining elements enhances knowledge retention. In this thesis, an educational game was developed to support young children (6-12 years old) to learn robotic coding. The study focused on robotics, an area that is important in everyday life and is expected to expand further in the future. The aim is to raise children's awareness of robotic coding and introduce basic programming concepts, create a platform for children to interact with the game and develop their basic skills such as coding, problem solving, algorithmic thinking and analytical reasoning. In an interactive learning environment designed for this purpose, functions have been created that allow children to design their own robots and understand coding principles through hands-on experience. The study also examined similar educational practices and analyzed their effectiveness in teaching robotic coding. The game was tested by 12 participants and their views on the game were evaluated. As a result of studies and researches, it has been observed that gamification positively affects children's learning processes. The developed game successfully combined educational content with entertainment, making robotic coding more accessible and engaging for young learners. This approach has contributed to an early interest in STEM fields and the development of cognitive skills. Future studies may explore further development and wider applications of game-based learning in robotic coding education.
Image processing applications in food industry
Traditional quality control methods in the food industry are mostly manual, leading to significant challenges such as variability of results, human error, and limited scalability. As the industry shifts towards automation and digitization, the potential of image processing as a solution for enhancing quality control processes becomes increasingly evident. This study is dedicated to applying deterministic image processing techniques, specifically those not involving machine learning algorithms, to streamline quality inspections in food production. The appeal of this approach lies in its simplicity, lower computational requirements, and ease of integration into existing production systems. Deterministic image processing, which employs defined algorithms for specific tasks, offers predictable and repeatable results. This thesis delves into techniques such as thresholding (including global and adaptive methods), edge detection (using operators like Sobel and Canny), and morphological processing to identify defects, contaminants, and structural inconsistencies in food products. Applying such methods ensures consistent quality checks and significantly reduces the reliance on human inspectors, thereby minimizing oversight and human error. For instance, thresholding allows for the segmentation of objects from the background, making it an essential tool for detecting surface anomalies and contamination. Edge detection methods facilitate the identification of product shapes and possible irregularities, contributing to more accurate quality assessments. Accordingly, no custom algorithm was written. Instead, reports were generated using Azure's Custom Vision API based on the captured images. The research methodology involves using Epson BT-350 smart glasses integrated with a custom-built Unity application for real-time image capture and analysis. Operators inspected pre-determined high-risk zones, collecting images processed using rule-based algorithms. These deterministic methods were chosen for their effectiveness in real-time applications, where processing speed and reliability are critical. The results from initial trials indicated that this system could accurately detect surface-level defects and contamination, with error rates reduced by up to 30% compared to manual inspections. This significant improvement not only highlights the potential for substantial cost savings but also underscores the enhanced operational efficiency that can be achieved through the use of deterministic image processing. One of the key findings of this research is the feasibility of implementing deterministic image processing methods in routine quality control without the need for complex machine learning models. While machine learning can offer adaptive and highly accurate solutions, it often requires extensive training data, higher computational resources, and specialized model development and maintenance expertise. In contrast, deterministic methods are more accessible and easier to implement and maintain, making them suitable for mid-sized and smaller food production facilities. This study also underscores the importance of understanding the specific challenges associated with image processing in a food production environment. Factors such as lighting variability, product movement, and diverse product types can affect the accuracy of automated inspections. By addressing these challenges and implementing adaptive techniques, such as local thresholding, which adjusts to changes in lighting conditions, the research ensures more robust performance, providing a comprehensive understanding of the complexities in the field. In conclusion, this thesis contributes to the field by demonstrating that non-machine learning-based image processing techniques can substantially benefit quality control in the food industry. These benefits include improved detection of defects and contamination, reduced human error, and enhanced operational efficiency. The system's adaptability and low computational overhead make it a promising and, importantly, practical and feasible solution for food manufacturers looking to modernize their quality control processes without significant infrastructure changes.
A Comprehensive Comparison of Open-Source MQTT Brokers Based on Architectural, Technical, and Functional Characteristics
daha sonra doldurulacaktır
Narwhal: Bulut altyapısını yönetmek için yeni bir görsel programlama dili
Modern software development processes have made it necessary to manage complex infrastructures and microservice architectures. This situation, especially with the popularization of Infrastructure as Code (IaC) approaches, has brought about the advantages of code-based solutions in infrastructure management, as well as increasing complexity and comprehensibility problems. This new visual programming language, called Narwhal, aims to provide an innovative solution to the current challenges of IaC by offering the possibility of performing infrastructure design through an interactive and visual diagram. Users can visually edit infrastructure components and configurations, convert them to Terraform HCL, and compile the infrastructure directly from this visual diagram. The visualization approach offered by Narwhal aims to reduce software architecture erosion by making infrastructure designs more understandable, shareable and easily manageable. The schematic structure, which allows for earlier detection of errors in software architectures, also significantly reduces documentation costs. In this study, Narwhal's contributions to infrastructure management, its potential to prevent software architecture erosion and its effects on modern software development processes are discussed in detail. Narwhal offers both a vision and a practical solution proposal for future infrastructure management tools.
Rag supported knowledge-based question-answer system with multimodal (text + visual) data
To address the inadequacy of traditional systems and Large Language Models (LLMs) in handling multimodal data (text, tables, figures) and complex PDF documents requiring privacy, this thesis presents a comprehensive and energy-efficient offline RAG (Retrieval-Augmented Generation) framework that overcomes existing online dependencies. The technical distinguishing feature of the work lies in its offline dual-channel multimodal architecture, which processes visual (Image) data through separate structures. This approach provides a closed and reliable solution by hosting Large Language Models (LLMs) locally on the device (Edge AI) without requiring an external API; this developed Hybrid System significantly outperforms the accuracy of basic VLM models (LLaVA and VILA's approximately 0.72-0.73 accuracy) in real-world tests, achieving an average success rate (CheckAvail score) of 0.8827. This proves it is a competitive alternative to the online, high-performance XSum architecture (0.97 score). The XSum score was found to be consistent with text-focused metrics. The Hybrid Score Model, which shows similar success to other hybrid models, particularly in evidence capture success (true0), unfortunately has the lowest value (0.69) in the metrics for excluding irrelevant evidence (true1); this indicates that the model lacks specificity and is prone to false positives (FP). Detailed analyses have revealed that this low performance in the true1 metric is primarily due to the evaluation of evidence from the visual (Image) channel, while the text channel produced satisfactory results. Additional analyses showed that these false positives were eliminated when verified using an offline visual LLM (Vision LLM). Work is ongoing to find a better solution in the future. Therefore, future work will focus on developing new-generation filtering mechanisms that will increase the model's specificity to overcome this problem, especially in the visual channel, architectural optimizations for resource-constrained LLMs, and robust image/visual extraction functions that can adapt to different PDF structures.
Dynamic integral sliding mode control of an electromechanical system
In this study, a new dynamic integral sliding mode control is put forward to control the speed of an electromechanical system. The dynamic integral sliding mode control is proposed to make sliding mode control aspects better and remove undesirable chattering effects. Lyapunov stability theorem is applied to guarantee the system's stability. The output speed, the sliding surface, and the control input of the electromechanical system for the controllers are examined. The simulation and experimental results obtained from the conventional sliding mode control are compared with the simulation and experimental results of the proposed dynamic integral sliding mode control under none, 20% and 80% of parameter uncertainty. The dynamic integral sliding mode control is much more robust than the conventional sliding mode control with respect to parameter variations.
Cloud coverage prediction with deep learning methods
In this work, cloud images classification and segmentation have been implemented by using Deep Learning techniques. All images utilized in this thesis were obtained from TUBITAK National Observatory Bakirliktepe Campus, Cukurova University Space Science and Solar Energy Research and Application Center, and Singapore Whole-sky Imaging CATegories database. The main goal of our study is to compare the traditional image processing results with the Deep learning techniques. Firstly, two classes' classification solutions have been used. Our goal was to separate the cloud images from the others. We highlighted on Convolutional Neural Network (CNN) as classification and on SoftmaxWithLoss for the prediction. During our training phase, 92% accuracy has been scored as after fine-tuning our model. Secondly, some image processing techniques have been used to detect and segment the cloud images. Some edge detectors and watershed techniques have been implemented. Thirdly, some segmentation solutions have been proposed. Fully Convolutional Network (FCN) and U-NET have been used in this thesis. Two different methods of Deep Learning have been proposed to segment the cloud images and then make the prediction. Using U-NET model for the segmentation, 87% as dice coefficient and 45% as loss have been scored during the testing step. Different learning techniques were implemented on FCN such as stochastic gradient descent, Adam's momentum technique and Nesterov's momentum technique. The highest segmentation accuracy on FCN has been 63.12% with Adam's momentum technique. Keywords: Deep Learning, Cloud segmentation, Convolutional Neural Networks, FCN, U-NET.
The effects of preprocessing methods on prediction of traffic accident severity
The purpose of this thesis is to investigate the effects of different preprocessing approaches on the prediction accuracy of classifiers regarding the severity of traffic accidents. For this aim, six different classification methods, including J48, Ibk, Random Forest, OneR, Naïve Bayes and SMO have been used on an imbalanced dataset consisting of 99% nonfatal and 1% fatal traffic accidents that took place in Adana between 2005 and 2015. Various undersampling and oversampling approaches are tried to solve the imbalance problem and improve the classification accuracy. Then, the results of each method are compared to determine the best classifier and preprocessing method. Accordingly, SMO has attained higher accuracy in nearly all analyses, and it has produced the highest scores with the undersampled dataset consisting of equal amount of nonfatal and fatal instances.
Investigation of single-rate triangular 3D mesh compression algorithms
In this thesis, currently available 3D mesh compression algorithms, frameworks, libraries etc. are investigated. Especially, the algorithms that are popular in survey papers but don't have any implementation or had outdated implementation or no published version is available, are gathered together and compiled accordingly. According to the benchmark test results, current best general-purpose data compression methods are identified and applied as the last stage of mesh compression. Results are compared in order to demonstrate the current state of single-rate 3D mesh compression performance with the current best general-purpose data compression methods.
Identity document image analysis using artificial intelligence techniques
Identity document understanding is one of the most important parts in the document recognition systems, even though that many researches are done in "document image analysis", however this research still has many challenges. This research aims to find a better solution to analyse the identity document image using artificial intelligence techniques. Hence, proposed method detects the location of the important information in the identity document, classifies the text to many categories (date of birth, last name, first name...etc.), and detects the key objects in the identity document image like the face photo and signature and even the logos. Methods used in this research are divivded into two categories which are document type based and machine-readable zone (MRZ) based approaches. In the document type based method, first the document image is classified, then pre-trained model for each class is utilized to derive the information needed such as location and text segments for that class on the query image. In the MRZ based approach, first MRZ location is detected, then Visual Inspection Zone (VIZ) area zone with the MRZ information are matched and small Natural Language Processing (NLP) engine is utilized to detect text. The accuracy of the research was 94.8% with accurate detection for all segments in the document, and 3% with good detection. However some information in the document image was missing, resulting in 1% of wrong detection as false positive, and the last 3 was not detected at all. Consequently, the overall accuracy of the research was 97.8% of the test samples.
A new deep learning approach: Differential convolutional neural network
Deep learning structures have achieved unprecedented success rates in many scientific research areas. The well-known deep structure, convolutional neural network, is commonly used in pattern recognition studies. Convolutional neural network structures are composed of a feature extractor consisting of convolution and pooling layers and a fully connected network used as a classifier. In the first study of the thesis, GCNN, a strong kernel-based classifier, was adapted to CNN structure to increase the classification performance. This adaptation led a relative performance increase up to 44.45%, 39.69% and 43.57% for precision, recall, and F1-score, respectively. Although this adaptation yielded a significant increase in performance, it was observed that the convolutional part was weak in terms of representation. Therefore, the idea of developing a convolution technique with higher learning performance has emerged. A novel convolution technique named as Differential Convolution which considers directional changes among a pixel and its neighbors is proposed. Deep structures applying Differential Convolution are named as Differential Convolutional Neural Networks. These structures made a relative performance boost up to 55.29%, 58.43%, 41.75% and 56.43% for accuracy, precision, recall, and F1-score, respectively. Key Words: Deep learning, convolutional neural network, convolution techniques, general regression neural network, image classification, pattern recognition, artificial intelligence, machine learning.
A semantic vector space model using Euclidean distance based relatedness
In this thesis, it is aimed to develop an efficient method to measure the semantic relatedness of the words. Although computer-based studies have achieved good results on this subject for nearly three decades, they have not succeeded to produce relatedness measurement close to human intuition. In this study, a WordNet-based approach is preferred, because WordNet has a graph adapted model. In addition, it is inspired by the so-called word embedding model, which is based on the dense representation of word prototypes in the low-dimensional vector space. Through proposed model, randomly positioned word prototypes are located into appropriate positions in multidimensional vector space with help of iterative learning algorithm that optimizes WordNet relation weights and word prototype positions that use Euclidean distance based relatedness. Both the positions of the words in the vector space and the weight of the semantic relations that connect the words on WordNet are determined effectively through the proposed model. The results obtained in the benchmark tests show that the new proposed model produces more successful results than the previous word-level semantic similarity studies. This approach might present a different perspective not only on semantic similarity studies but also on solving many other natural language problems.
Development of a new software for fabric defect detection and classification using image processing and machine learning methods
Quality control in the fabric industry involves a set of standards or guidelines that help guarantee a product meets certain parameters as well as customer satisfaction. Fabric defect detection (also called inspection) is a quality control process aimed at identifying and locating defects. The aim of this thesis is to build an application based on image processing and deep learning methods to automatically detect the defects on the fabric surface and classify them. Discrete Fourier transform (DFT). Normalized cross-correlations, homogeneity equalization, and Gabor filters have been used in image processing, Faster Region Proposal Networks (Faster-R CNN) has been used in classification. The mean square errors (RMSE's) has been used for computing the detecting and classification errors. Based on the results obtained, it has been proved that image processing methods especially DFT can be used for the defect detection with acceptable results.
Modified recurrent convolutional neural networks for action recognition
In this study, two new neural networks are proposed for action recognition. The first method is named as modified recurrent CNN and the second method is named as feature concatenating CNN. The main aim for proposing these networks is developing new techniques for sharing features between different frames of the videos. These methods have been applied on different datasets. Those datasets are UCF101, Hollywood2 and HMDB51. Based on the results, modified recurrent CNN is a good alternative for facilitating better feature learning. In most of the cases, accuracies provided by the modified recurrent neural networks are a few percent larger than accuracies provided by the standard convolutional neural networks. Additionally, a detailed review of the most important action recognition methods that are based on deep learning has been provided.
Privacy preserving rule-based classifiers using modified artificial bee colony optimization algorithm
Privacy preserving data mining is a hot research field for data mining. The aim of privacy preserving data mining is to prevent the leakage of the sensitive information of individuals while performing data mining techniques. Classification task is one of the most studied fields in data mining hence in privacy preserving data mining as well. On the other hand, differential privacy is a powerful privacy guarantee that determines privacy leakage ratio by using ϵ parameter and enables researchers to mine data which includes sensitive information. Although the success of the rule-based classifiers using meta-heuristics such as Ant-Miner etc. in data mining has been demonstrated, any implementation of these classification algorithms with differential privacy has not been proposed in the literature to our best knowledge. Motivated by this, implementations of the rule-based classification algorithms by using meta-heuristics with differential privacy are performed in this thesis. According to the experimental results, the proposed rule-based classification algorithms outperform the other classification techniques in the literature for low ϵ parameters (i.e., ϵ=1). Key Words: Differentially Private Rule-Based Classifiers, Artificial Bee Colony Optimization, Privacy Preserving Classification.
Cyberbullying detection using text classification for turkish language
Cyberbullying is an electronic form of peer harassment. It includes relational attack behaviors such as harassing people, mocking people, threatening, spreading gossip, and insulting people on the internet by using information and communication technologies. In Turkey and many European countries, the cyberbullying is considered as a serious problem after the cyberbullying related suicides occurred. In recent years, researches are being carried out and solutions are tried to be found by experts, especially with educational scientists and psychologists, about cyberbullying. The aim of this study is to create the largest Turkish dataset so far for the detection of cyberbullying texts and to show the effects of preprocessing, feature selection and classifiers for the detection of cyberbullying from texts. In this study, a number of preprocessing steps are applied, and two well-known filter-based methods that are information gain and chi square are used for feature selection. Among the classifiers tested, Naive Bayes Multinomial is determined to be the most successful method for detecting cyberbullying from texts written in Turkish language. In addition, a filter-based classifier is proposed, and its performance is tested on the collected dataset. The proposed method has promising accuracy and can be used for labeling any Turkish text document without re-training the classifier.
Application of particle swarm optimization for computer aided diagnosis of diseases
Data mining is used in order to obtain meaningful information from the data obtained in many different fields by applying several methods. Data mining is widely used to analyze medical data to make diagnosis of several diseases as this is very important topic and there exists large amount of available data in medical domain. In this study, Particle Swarm Optimization (PSO) is used to reduce the size of the medical data by making feature selection to perform better data analysis from healthcare datasets. To reach our goal, Breast Cancer Coimbra, Diabetic Retinopathy Debrecen, Self-Care Activities, and Lee Silverman Voice Treatment datasets that are used to diagnose breast cancer, diabetic retinopathy, children's self-care problems, speech disorders of patients having Parkinson disease, respectively, obtained from UCI Machine Learning Repository, are analyzed by using Naïve Bayes (NB), Support Vector Machines (SVM), and Random Forests (RF) classifiers. The experimental analysis has shown that the PSO based feature selection improves the classification accuracy of diagnosis of diseases for the NB and RF classifiers. The PSO based method is also compared with the well-known feature selectors that are information gain (IG), chi-square (CHI2) and Relief. It is observed that PSO based method has better performance than that of IG and CHI2 methods, and similar results with the Relief method.
Predicting maximal oxygen uptake using deep learning
Maximal oxygen uptake, or VO2max, is an external parameter that is affected by things like how many red blood cells the body has, how adapted the muscles are to distance running, and how much blood the heart can pump. It is measured as milliliters of oxygen used in one minute per kilogram of body weight. In a laboratory, it is calculated by measuring the volume (V) of oxygen (O2) that the body consumes while running on a treadmill which is the most accurate way. However, because of the serious drawbacks of direct measurement, a lot of studies have been conducted using machine learning methods to predict VO2max. The purpose of this study is to build new VO2max prediction models using deep learning (DL). The dataset has been split into training and test data using 70-30%, 80-20% split ratio, and 10-fold cross-validation. For comparison purposes, VO2max prediction models based on Multi-Layer Perceptron (MLP), Support Vector Machine (SVM), and Single Decision Tree (SDT) have also been developed. The performance of the prediction models has been evaluated using Standard Error of Estimate (SEE) and Multiple Correlation Coefficient (R). As a conclusion, DL can be used safely in VO2max prediction domain.
A study to improve performance of genetic algorithms
Selection is one of the most crucial steps of Genetic Algorithms (GAs) that commonly used in areas of robot applications, image and voice recognition, artificial intelligence applications, path finding problems, scheduling problems, etc. In GAs, lack of adjusting the balance between exploration and exploitation, and selecting appropriate parameter settings are main problems of most selection methods as they cause premature convergence and trapping in local optima. In order to overcome these problems, two common techniques have been utilized: presenting a new selection method, tuning the parameter of an existing algorithm. In the first part of study, new selection methods, Aggressive, Non-Aggressive, Integrated Aggressive, Integrated Non-Aggressive, Outlander, Non-Aggressive Outlander, Bipolar Mating Tendency (BMT), were proposed to solve the problems. As most of the methods are based on Standard Tournament Selection (ST), their performances were compared with ST and prevalent selection methods that are also based on ST: Restricted Tournament, Unbiased Tournament, Fine-Grained Tournament and Cooperative Selections. Twenty-one well known test functions in the field of GAs were employed for the comparison. Furthermore, non-parametric statistical tests, Friedman and Wilcoxon Signed Rank, were applied to demonstrate the significance of the results. In the second part of the study, meta search methods (Brute Force and Coarse to Fine) and meta optimization algorithms (GAs, Particle Swarm Optimization and BMT) were applied to tune standard GAs in order to achieve its best performance. Moreover, the second part contains a short survey in the literature of meta optimization.
Development of a new software for resting metabolic rate prediction using machine learning methods
Resting metabolic rate(RMR) indicates the number of calories that are needed to carry out basic functions in mammals like blood circulation, brain functions, breathing, fuel ventilation, temperature regulation, etc. at complete rest. Accurate prediction of RMR plays a critical role to detect an individual's establish daily calorie needs, risk of heart disease and stroke, hypertension, diabetes, and internal age so nutritionists calculate RMR of patients within the specified period and able to prepare diet lists for them to help their wellness goals. In this thesis, it was objected to develop a novel web-based application that can predict the individual's RMR using different machine learning methods. The application has been developed using DTREG predictive modeling software library, Visual Studio, and C# programming language. Three different machine learning methods which are General Regression Neural Network (GRNN), Multi-Layer Perceptron (MLP), Support Vector Machines (SVM) have been integrated into the software. Different prediction models have been assessed according to their Root Mean Square Error (RMSE) metrics. As a result, it has been proven that this software can be used for RMR prediction, producing acceptable error rates under certain circumstances.
Gait-based gender classification using neutral and non-neutral gait sequences
A biometric system provides automatic identification of an individual based on his/ her unique feature or characteristic. Biometric identifiers are often categorized as physiological versus behavioural characteristics. Physiological characteristics are related to the shape of the body, like fingerprint, palm veins, face shape, while behavioural characteristics are related to the pattern of an individual‟s behaviour, including gait, handwritten signature and voice. Gait as a means of biometric recognition aims to recognize a person by he/she walks. Human gait feature could be used in different applications such as identifying unauthorized persons, identifying their gender, and determining walking-related abnormalities by analysing the way they walk or move. In this thesis we aim to propose gender classification based on human gait features to investigate the problem of non-neutral gait sequences: coat wearing and carrying bag conditions in addition to the neutral gait sequences. Our objectives will focus on investigating and testing the performance of gait sequence features for the purpose of gender classification. Our tests are based on large number of experiments using CASIA B gait database, includes 124 subjects (31 women and 93 men), recorded from 11 different view angles. For each subject, there are 10 walking sequences, consisting of 6 Neutral sequences (Nu), 2 Bag-Carrying sequences (CB) and 2 Coat-Wearing sequences (CW). The proposed method mainly classified into three parts; the first part is focused on investigation of isolating conductive frames from their backgrounds using the frame differencing method. The second part is related to feature extraction, for which we propose a new set of features which are constructed as based on the Gait Energy Image and Gait Entropy Image, called Gait Entropy Energy Image (GEnEI). Three different feature sets are structured from GEnEI based Wavelet Transform, called Approximation coefficient Gait Entropy Energy Image (AGEnEI), Vertical coefficient Gait Entropy Energy Image (VGEnEI), and Approximation and Vertical coefficients Gait Entropy Energy Image (AVGEnEI). Finally two different classification methods are applied to test the performance of the proposed methods separately, called k-Nearest-Neighbour (k-NN) and Support Vector Machine (SVM). Further, these three sets of features are tested separately using the fused-based decision level fusion method. We demonstrate that when k-NN is used as a classification method, AGEnEI results in 97% fusion level for Nu gait sequence, VGEnEI results in 91.4% fusion level for CB sequence and for CW sequence AGEnEI produces 83.6% fusion level. Among three sets of features, k=1 notably produces better average fusion level compared to the other two sets of features, i.e. k=3 and k=5. When three sets of features (AGEnEI, VGEnEI , AVGEnEI ) are fused using the decision level fusion method, we obtain accuracy of 99.8%, 92.2% and 86.3% for Nu, CB and CW respectively. These results outperform the results achieved when each of these sets of features are applied separately.
Face recognition system based on PCA-wavelet and support vector machines
Face recognition can represent a key requirement in various types of applications such as human-computer interface, monitoring systems as well as personals identification etc. In this thesis, different types of methods were used for implementing and testing face recognition system. These methods are first way introduced in one face recognition system and the combination of these methods together support the recognition system and give better results compared to other methods. In the first part, a combination of PCA and Wavelet Feature Extraction methods are used to obtain the important features and reduce the dimensions of the face image. In the second part, SVM classifier is used to classify the features of image and K-Nearest Neighborhood is applied to identify it. In the third part, Classification performance under different kernel types is examined to demonstrate the classification performance of SVM, in addition, to compare the SVM classifier result with artificial neural network classifier. Finally, the performances of the system under various conditions are tested. For a more comprehensive comparison, two face image databases are used to test the performance of the system. The experimental results proved the efficiency and reliability of the system and the results enhancement of 5% by using the SVM classifier with polynomial Kernel Function compared to use feed forward Backpropagation neural network classifier.
Automatic synset detection from Turkish dictinary using confidence indexing
Bu çalışmada, bir Türkçe anlamsal ağı, bilgisayar okunabilirliği olmayan tek dilli sözlükten tasarlanmıştır. Sözlük madde başları ve tanımları ağırlıklı iki parçalı çizge modeline işlenmiş ve anlamsal ilişkiler açısından analiz edilmiştir. Genel anlamsal ağının üst anlamlı, eş anlamlı ve karşıt anlamlı olarak birincil anlamsal ilişkileri Madde Başı-Anlam sözlüğüne göre analiz edilmiş ve anlam düzeyinde anlamsal ağa eklenmiştir. Eş anlamlı ilişkiler, geliştirilmiş bir eş anlamlılar kümesi tespiti için bir güven seviyesi ile etiketlenir. Ayrıca, madde başları ile bu madde başlarından oluşturulmuş olan türemiş ve bileşik madde başları arasında biçim-anlamsal ilişkiler eklenmiştir. Ayrıca N-Gram analizi, herhangi bir ek anlamsal ilişkinin örüntülerini bulmak için kullanılmış ve örüntüleri bulunan ek anlamsal ilişkiler, anlamsal ağa eklenmiştir. Son olarak, eşanlamlılar, kapsayan ağaç tabanlı eş anlamlılar kümesi algılama algoritması ile eş anlamlılar kümesi oluşturmak için kümelenmiştir. Elde edilen eş anlamlılar kümesi, güncel ve kapsamlı bir Türkçe wordnet ile karşılaştırılmıştır.
Investigation of multi-focus image fusion
In this study, a study has been carried out on the multi-focus image fusion that is the sub-branch of the image processing area. The aim of this thesis is to propose a new and effective multi-focus image fusion method that takes advantage of the successful aspects of present methods and improves the missing aspects. The goal of multi-focus image fusion is to provide a completely focused image by distinguishing between focused and unfocused pixels. Based on the literature review, it has been noticed that the importance of edge information in distinguishing pixels is emphasized. It has been concluded that the difference between focused/ non-focused pixels can be revealed more clearly with gradient images rather than the original images. The energy of Gradient (EOG) and standard deviation functions have been used as the focus measurement on gradient images. In order to minimize the error, the majority voting method is applied with different window sizes for the final fusion. The proposed method has been compared with seventeen different new and traditional multi-focus image fusion techniques both visually and objectively. Six different quantitative metrics have been used for objective evaluation. It has been observed that the proposed method is promising according to visual evaluation and 83.3% success has been achieved by being first in five out of six metrics according to objective evaluation.
Performance analysis of congestion control protocols in internet of things networks
The extensive flow of information in the Internet of Things (IoT) networks greatly increases the importance of congestion control. Therefore, in the academic literature, there are many studies about how to solve congestion control and on which layer to handle it. In this thesis, it is examined how the protocols written for congestion control on the application layer (CoAP, CoCoA, etc.) perform experimentally by running various combinations of several sub-layers with different protocols in IoT networks. In addition, the relationship between the application layer protocols and the objective functions of the network layer protocol RPL is examined. As a result, it has been proved that OF0 performs better than MRHOF, and CoCoA Strong was observed as the best performing congestion control mechanism. The most efficient results were obtained in scenarios where OF0 was used with ContikiMAC.
Prediction of electricity market clearing price using machine learning and deep learning
The market clearing price is the equilibrium monetary value of a traded asset or good and is an important metric in the calculation of electricity prices. In our country, when the electricity price is determined by the companies and in the price changes, market clearing prices are calculated. In this study, forecasting models consisting of 24 forecasts were created in order to make one day forecasts using hourly market clearing price data. The main objective of this thesis is to estimate the market clearing price of electricity using Multilayer Perceptron (MLP), Recurrent Neural Network (RNN), Convolutional Neural Network (CNN) and Long Short Term Memory (LSTM), which are machine learning and deep learning methods and to create appropriate models. The performance of the predicted models was evaluated by calculating the Mean Absolute Percent Error (MAPE) value. The results generally show that LSTM (Long Short Term Memory) and CNN (Convolutional Neural Network) based models perform better than other methods. Keywords: Machine and Deep Learning, Forecast, Market Clearing Price
Exploring mini-batch sample selection strategies for deep learning based speech recognition
This thesis aims to propose mini-batch sample selection strategies for deep learning based speech recognition systems. Deep learning based speech recognition systems became more prevalent and state-of-the-art system for speech recognition domain with the popularity and success of deep learning architectures. Recurrent Neural Networks (RNN) and Long Short-Term Memory (LSTM) RNN are widely and successfully utilized for the applications of speech recognition. Mini-batch gradient descent algorithm is generally accepted algorithm for training deep learning based speech recognition systems. Mini-batch gradient descent algorithm is a successful algorithm, but one of the main problems in mini-batch gradient descent is that the training samples are selected randomly for each mini-batch. In this thesis, mini-batch sample selection strategies are proposed to improve speech recognition accuracy of deep learning based speech recognition systems. Proposed strategies use meta features of speech corpuses, i.e. gender and accent features. Three types of sample selection strategies are proposed, i.e. gender adjusted strategies, accent adjusted strategies, and hybrid strategies that combine gender and accent adjusted strategies. The experimental results show that proposed strategies are beneficial for improving performance of deep learning based speech recognition systems.
Applications of the pressure effect on the electrical energy: Piezoelectric example
Energy needs have increased due to the rapid development of technology from the past to the present, rapid population growth and associated production increase. Fossil fuels are generally used to meet this need. In recent years in all countries; due to the facts that fossil fuels will be depleted and the use of these fuels harms nature, a vast amount of resources are spared for researches of alternative energy sources. The facts that they do not harm the nature, they will not be exhausted in the future, they use various sources (solar, wind, water, geothermal...), and that there is one or more alternatives appropriate for each country have increased interest in renewable energy sources. In this study; the transformation of the movements of the living things (basically human) into electrical energy using piezoelectric materials was investigated. For this purpose, floor of 1 m2, which can harvest was formed. The energy harvesting floor is designed using low-cost materials. In the floor design, piezodiscs with thicknesses of 0.33 mm and 1mm were used separately to analyze the floor. For the floor formed with piezodiscs of 0.33 mm thickness, $ 28.58 was spent and a maximum of 3,976 mW was produced as a result of the walking of a person weighing 60 kg for 1 minute. Approximately $ 50.16 was spent for the floor created with 1mm thick piezodiscs, and 4,029 mW was produced as a result of the walking of a person weighing 60 kg on it for 1 minute. The result of the experiment is that the current produced is very small, and that the costs of the renewal of damaged materials, initial installation and operating costs are quite high. It was observed that the cost of storage of the energy generated from the floor does not have economic value when coupled with production cost, material and operating costs. At the same time, the energy produced does not have the power to be used. In a much larger area, with a greater number of people (Shopping entrances, factory establishments...) a certain level of energy can be obtained but it will not be efficient. Therefore, it is not possible to operate with piezoelectric materials by establishing an efficient clean energy system with the design in the study or with similar designs. The different design works need to be sustained in order to use the materials in the industrial area. It is believed that the thesis study will be the source of further studies in this field. For this purpose, all the design steps are shown, the materials are introduced and the technical information is included faultlessly in the study.
Music emotion recognition using convolutional long short term memory deep neural networks
In this thesis, we propose an approach for Turkish music emotion recognition based on convolutional long-short term memory deep neural network (CLDNN) architecture. For this purpose, a new Turkish emotional music database composed of 124 Turkish traditional music excerpts with a duration of 30 seconds each is constructed. We used novel features obtained by feeding convolutional neural network (CNN) layers with log-mel filterbank energies and mel frequency cepstral coefficients (MFCC) in addition to standard acoustic features. Classification results show that the best performance is obtained when the new feature set is combined with the standard features using the LSTM + DNN (LDNN) classifier. The overall accuracy of %99.19 is obtained using the proposed system with 10-fold cross-validation. When new features are added to the standard features, 6.45 and 5.65 points improvements are achieved for native listeners and experts, respectively. Additionally, the results also show that the LDNN classifier yields 1.61, 1.61, 2.42 and 3.23 points improvements for native listeners and 5.65, 0.81, 4.84, and 6.45 points improvements for experts in music emotion recognition accuracies compared to that of K nearest neighbors (k-NN), Sequential Minimal Optimization (SMO), Naïve Bayes and Random Forest (RF) classifiers, respectively.
Performance analysis of deep learning object detection based image segmentation methods
Image segmentation which divides the input image into multiple regions or segments is one of the most difficult problems to be solved in the area of computer vision. Without image segmentation, understanding the images by computer is not an easy process. There are two types of image segmentation task: i) semantic segmentation, in which multiple objects from the same class are considered as the same, ii) instance segmentation where multiple objects from the same class are taken as different, therefore, each object (instance) is to be classified separately. Image segmentation is used in numerous applications such as satellite image processing, medical image processing, texture recognition, face recognition systems, automated plate recognition systems, etc. In this thesis, our aim is to apply deep learning-based object detection to perform semantic segmentation and evaluate its performance. Therefore, first, we applied YOLO to find the bounding boxes of each object in the images, then we used GrabCut and DeepGrabCut methods to classify foreground and background pixels to make semantic segmentation. DeepGrabCut is a deep learning-based version of GrabCut, and we compared their performances on two image datasets having more than 35K images. The experimental analysis has shown that object detection with object selection can be used to make semantic segmentation with acceptable error rates when optimal parameters setting was done for the methods applied. Keywords: Deep learning, prediction, image segmentation, object detection, computer vision
A comparative study of deep learning methods for classification of rna-seq cancer data
Cancer is one of the most important causes of deaths today. Millions of people die because of cancer every year, while millions of people are diagnosed with cancer. Cancer is a gene disease. As a result of mutations in genes, cells become abnormal and uncontrolled division is the main cause of cancer disease. Therefore, gene expression is very important in the diagnosis and classification of cancer. RNA-Seq data stores information of many genes. Many of these genes found on RNA-Seq data have nothing to do with cancer. Finding which genes cause cancer and then diagnosing the type of cancer is a long time process. Decision support systems can be developed using classification algorithms or deep learning methods to shorten this process and assist doctors in the diagnosis process. The aim of this thesis is to analyze the cancer type using clasical methods, artificial neural networks and deep learning methods by using RNA-Seq datasets created with genes obtained from previously diagnosed cancer patients. First, gene selection is made using wrapper methods to reduce the size of the RNA-Seq data set. The selected genes are then used in the classification process. For classification, decision trees, random forests, support vector machines, artificial neural networks and deep learning methods are used. After this study, which method works better in cancer classifications is examined. The method developed according to the results is expected to help doctors in the process of cancer classification. Keywords: Cancer, Gene Expression, RNA-Seq, Classification, Deep Learning
Social media text classification for crisis management
In recent years, impressive attention has been given for mining the publically available huge amount of data to gain situational awareness, which may help in preventing or decrease the effect of some disaster by taking the correct responses. In this study, an effective Convolutional Neural Networks (CNN) tweet classification system that fully supports the Turkish language has been developed. In addition, the first-ever Turkish tweet dataset for crisis response is created. This dataset has been carefully preprocessed, annotated, well organized and suitable to be used by all the well-known natural language processing tools. Furthermore, the performance of some well-known machine learning algorithms, i.e., K-Nearest Neighbor (KNN), Naive Bayes (NB), and Support Vector Machine(SVM) was investigated. Then, the performances of the ensemble systems Random Forest (RF), AdaBoost Classifier (AdaBoost), GradientBoosting Classifier (GBC), when used for text (tweets) classification, has been also observed. A wide range of experiments was performed to investigate the performance of the developed system. As a result, the developed approach has achieved very good performance, robustness, and stability when processing both Turkish and English languages. Key Words: Crises Management Systems; Tweet Classification; Turkish language; Convolutional Neural Networks; Natural Language Processing.
A comparative study of deep learning methods on flexural buckling load prediction of aluminum alloy columns
In recent years, aluminum alloy columns have been widely used in construction fields. This is due to the light weight of aluminum alloys, high corrosion resistance, long life, low maintenance costs, the possibility of recovery, versatility of the metal and the possibility to obtain endless variety of profiles has many advantages. The calculation of the critical buckling loads of the columns is the most important issue. However, it is known that heat treated aluminum alloys have higher proof stress yield strength than non-heat treated aluminum alloys. In this study, buckling load estimation of heat treated aluminum alloy columns is made by using deep learning method and soft computing techniques and laboratory test results are compared. Sequential Model is used while using deep learning method. Adam, Adamax, Nadam, Adadelta, Adagrad, RMSProp and SGD and the optimizer functions of deep learning are evaluated separately. In addition, the results are evaluated using both MAE and MSE Loss functions for each optimizer. As a result of the study, it is understood that the optimizer and loss functions used together are more successful when estimating value for the dataset using deep learning model.
Attribute inference over real-world online social networks: a comprehensive privacy analysis
Today, people can make interactions through online social networks (OSNs) that allow users to reveal and see their personal information, make connections with other users, and view the public information of other users through the connections. As such, OSNs are extremely popular and inevitable parts of people's daily lives today. As a result, OSNs have become attractive data sources for researchers from many disciplines including computer science to investigate privacy, security, and user behavior issues. However, studies that aim to make privacy analysis of Turkish OSN users are quite limited. In this thesis, a comprehensive privacy analysis of Turkish Facebook users was carried out due to the popularity of Facebook OSN. Different methods were used to infer users' private attributes including gender, kinship, and so on. Then the inferred attributes were used in the privacy risk analysis to show that users are at higher risk than they believe. Additionally, two natural language processing (NLP) tasks including sentiment analysis and named entity recognition have been performed as they can be utilized in privacy risk analysis. NLP and feature inference tasks were handled as sub-problems. In these tasks, various contributions were made to the literature, and some of them provided state-of-the-art results.
A Turkish broadcast news speech database for investigation of the effect of deep neural network and long short term memory hyperparameters on speech recognition based systems
Speech recognition is the transformation of spoken words and sentences into text. It is used for various dictation operations as well as voice control applications. There have been many studies on speech recognition in many countries recently, the biggest reason being that they have large speech datasets in their own language and are accessible to them. However, studies on speech recognition applications in our country are very few, one of the reasons is the lack of voice dataset. In this study, a Turkish speech database has been developed for Turkish speech recognition based systems. Sound recordings were obtained from news broadcasted by Turkish News Tv Channels at different times. The stages of database creation are examined step by step and the tools used are discussed. The created dataset was shared on the web in a way that everyone can access in order to set a precedent for other studies. Additionally, we investigated and compared the effect of the number of layers and number of cells hyperparameters on Long Short Term Memory (LSTM) and Deep Neural Network (DNN) models on Turkish Broadcast News Speech Dataset that we created.
Fully automated deep learning and machine learning –based prognosis models for survival prediction of brain tumor patients using multi-modal mri images
Brain tumor is one of the most deadly types of cancer diseases. Accurate assessment of pre-surgical prognosis for patients with this disease can lead to better patient management. While Biopsy is the most commonly used diagnostic technique in routine clinical applications of prognosis estimation, it has several disadvantages such as it is invasive, and prone to tissue trauma. Consequently, automated pre-operative prognosis estimation techniques based on MRI images are recently getting attention, so noninvasive. However, most of the recently developed automated techniques are based on the handcrafted image features extracted from the manually segmented tumor regions in MRI, which is tedious & time-consuming. This study aimed to develop fully automated pre-operative prognostic models for the survival time, and glioma grade predictions in multi-modal MRI images of patients with brain tumors by using two-stage learning-based methods. In the first stage, we developed novel CNN architectures using pre-trained deep learning models as backend. In the second stage, the outputs of CNN models were fused using various classical machine learning methods to get the final prediction results. The experimental results demonstrate that the proposed prognostic models achieve AUC values of 99.7%, and 93% in glioma grading, and survival time predictions, respectively, outperforming current state-of-the-art results.
Predicting the performance of cross-country skiers using maching learning methods
The purpose of this thesis is to develop new regular and feature selection-based models for predicting the racing times of cross-country skiers by using machine learning and feature selection methods. Particularly, six popular machine learning methods including Optimized-General Regression Neural Network (OPGRNN), General Regression Neural Network (GRNN), Support Vector Machine (SVM), Multilayer Perceptron (MLP), Radial Basis Function Neural Network (RBFNN), and Single Decision Tree (SDT) have been used, whereas Relief-F has been employed as the feature selector. Several models have been developed to predict the racing time of cross-country skiers using physiological data along with a rich set of survey-based data. By performing 10-fold cross-validation, the prediction errors of the models have been calculated using root mean square error (RMSE). The results emphasize that OPGRNN-based prediction models show superior perfor¬mance and can be categorized as a feasible tool to predict the racing time of cross-country skiers. Furthermore, significant advantages such as the non-exercise-based usage and the applicability to a broader range of cross-country skiers make the prediction models proposed in this study easy-to-use and more valuable. Key Words: Machine learning, racing time, cross-country skiers, prediction
LGB tabanli toprak tekstür anali̇zleri̇nde test süresi̇ni̇n kisaltilmasi
In this thesis, an approach based on curve fitting, support vector regression, multilayer perceptron and long short-term memory architecture is proposed to shorten the experiment time in experiments with the Laser Guided Bouyoucos device developed for soil texture analysis. For this purpose, texture analysis signals obtained from 52 soil samples, each with 14400 samples (2 hours), were used. In the traditionally used curve fitting method, the shortest signal segment is found with an acceptable absolute error by shortening the soil texture signals from the end, while in machine learning methods, the shortest signal segment is found starting from the beginning. In the curve fitting method, the most suitable curve for the soil signals was selected as the 2nd degree exponential equation with the R-squared method. In the time shortening study with SVR, the model was trained and tested in the sample range of 1000-7000. In order to determine the entrance segment length in the MLP method, tests were carried out with a segment size in the range of 50-950 and the entrance segment size was selected as 200. In the MLP method, 3 layers are used: 1 input layer with 200 inputs, 1 hidden layer with 100 neurons, and 1 output layer with output. In the LSTM method, a 3-layer architecture is used, including the 1-input input layer, the 200-neuron LSTM layer, and the single-output output layer. As a result of the time shortening studies, the soil components were estimated with the SVR.
Design of a soil texture analysis device based on ultrasound sensors and machine learning methods
In this thesis, a digital soil texture analysis system is designed and introduced, which can be an alternative to the traditional hydrometer method used to find the proportional distributions of sand, silt and clay minerals in the soil. Traditional methods have many disadvantages such as being completely mechanical, needing expert control and laboratory. Considering today's advanced technologies and innovations, it has become inevitable to design a computerized forecasting system. The system has been redesigned using a 3D-printed container with ultrasound sensors. The system makes predictions by interpreting the changes in the intensity of the sound signals passed through the soil-water mixture placed in a closed container with machine learning methods. The changes in these signals, which are obtained by utilizing the sedimentation properties of sand, silt and clay particles in the soil-water mixture at different rates, were recorded on the computer, and computerized estimation steps were applied to the data. By using Support Vector Regression and Multi-Layer Perceptron architectures, the success of machine learning methods have been compared against traditional hydrometer results of the sample soils. Considering the 10% margin of error accepted in the standard hydrometer method, it has been seen that the proposed machine learning supported automated texture analyzer produced acceptable results. Thus, a computerized soil texture analyzer, which can predict the percentages of sand, silt and clay in the soil-water mixture in a closed container, using machine learning methods, is independent of expert supervision and laboratory environment, has a high portability, and can work with less material, has been presented in detail.
Fault-tolerant sliding mode control design for an electromechanical system
The purpose of this thesis is to propose an adaptive dynamic proportional–integral–derivative sliding surface based second-order fault-tolerant sliding mode controller for speed control of an electromechanical system under uncertainties and disturbances. The sliding-mode control technique is a robust control scheme that obtains the desired output with a concept of changing the controller's structure in response to change the state of the system. The online adaption skill of adaptive control makes it superior to sliding mode control in terms of constant or slowly-varying parameters for a non-linear dynamical system having uncertainties. Fault-tolerant control schemes are utilized commonly in safety-critical systems. This study aims to show how to utilize the robust properties of second-order adaptive dynamic sliding-mode control on the problem of component fault. The control design process prioritizes, even in the existence of the unsuitable conditions, providing the closed-loop stability for overall system. The Lyapunov theorem is used for affirming the design and stability of the closed-loop system. The presented adaptive dynamic first and second-order fault-tolerant sliding-mode controllers and the traditional sliding-mode controller are compared with each other and the results are discussed. The experimental results of the proposed fault-tolerant controller, concerning parametric uncertainties and disturbances, acquire suitable tracking performance and show more robustness than the traditional sliding-mode control.
Depth from blur
The most accessible and appropriate approach to recording and storing the depth measurements collected from a scene is through a depth map; accurate depth maps are also essential in extended reality and movie production. This topic of study is both intriguing and beneficial, and lucrative. Several firms are developing depth estimates for a range of reasons. Some may use it to add effects (such as bokeh) to photos and selfies (portrait mode) based on distance from the camera; others, both aerial and terrestrial, may utilize it for replacing or supplementing existing sensors in autonomous vehicles. A depth map is a two-dimensional array with the x and y distance information corresponding to the array's rows and columns, as in a conventional picture. This study examines the single-image depth inference problem using focus and blur images. A comparative work that examines Carvalho's, Lee's and Laina's methods that produce a depth map from a single image is carried out in this thesis. Carvalho's method takes a single synthetic to defocus image as input, and the output is the depth map using D3-Net. Both Lee's and Laina's approaches build a depth map from a single image using encoder-decoder architecture and Residual Network (ResNet50), respectively. After that, the predicted depth maps were segmented into three classes (near, far, far away). This study aims to determine the best performing method for estimating the depth map using the New York University v2 dataset. Furthermore, we can use the segmentation results to navigate the cameras (drones, robots, autonomous vehicles, etc.). As the results show, Carvalho's method was the best in-depth map estimation because of the synthetic defocus image dataset. Nevertheless, Laina's method is the best segmentation in near and far areas. The experimental results demonstrate that the NYU v2 dataset used with these models achieved accuracy values of 99.8%, 99.0%, and 98.8% for Carvalho, Laina, and Lee, respectively, for the predicted depth map. For segmentation, the accuracy values were 77%, 55%, and 90% for Carvalho, Lee, and Laina, respectively. Keywords: Depth map, Blur, Recurrent Neural Network, Convolutional Neural Network, Deep Learning
A modified autoencoder approach for feature selection
As the technology improves, data sizes have become huge. This also brought difficulties in extraction of meaningful information. As a result, new data analysis methods have emerged. Since data collection is everywhere in our daily life, data includes many redundant and unnecessary records and features. To identify useful part of data, feature selection algorithms have been used for a long time. However, those algorithms should be improved to deal with large scale data. In this thesis, we developed a new autoencoder based feature selection algorithm. Unlike traditional use of autoencoder, in this study, trained weight values are utilized instead of transformed data. The main idea behind the method is if the average weight of an input is high, it should be a useful feature. This simple but effective method was tested on 5 different datasets. 4 of them are standard datasets from Kaggle and UCI repositories. One of them is drug-target prediction dataset which is very difficult to classify due to imbalance nature of the data. While proposed method provided good results on standard datasets, not only proposed method but also all other tested methods provided very low results on drug-target interaction dataset due to the imbalanced nature of the dataset. Key Words: Machine learning, Deep learning, feature selection, drug-target interaction, autoencoders
Comparison of RNN-CTC, LSTM-CTC and GRU-CTC models and parameters on a new Turkish audiobook dataset
Speech is very important in human communication. Speech recognition systems work to convert sounds and text. Devices that use speech recognition systems make daily life easier. Although there are many studies on Turkish speech recognition systems, the lack of data sets is obvious. In this thesis, the original Turkish Audiobook Dataset was developed and neural network models were examined. An original data set obtained from audiobook recordings was prepared. Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Long Short Term Memory (LSTM), Gated Recurrent Units (GRU), Connectionist Temporal Classification (CTC) models were examined and compared on the obtained data set.
Anomaly detection in network traffic using machine learning
A primary thematic of this study is centered on detecting anomalies and measuring the device health for Central Processing Unit (CPU), memory utilization, and allocation; for Key Performance Indicator (KPI) dataset which assembled throw twenty-one-day, by improving models using machine learning (ML) methods; namely, Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM), with Auto Encoders (AE), One-Class Support Vector Machine (Oc-SVM), also k-Nearest Neighbors (k-NN). The accuracy of all methods was measured by using a confusion matrix. According to the observed results, the deep learning methods yield great performance results compared to classification methods for all models. In general, CNN/AE and LSTM/AE models show higher accuracy than the other methods. The ranking of models from best to worst based on accuracy in the confusion matrix are; CNN/AE, LSTM/AE, as for the deep learning models, while for classification models the favorable order for the methods are; k-NN, and Oc-SVM.
Development of recommender system algorithms for cold-start problem
Cold-start problems are one of the most important challenges in recommendation systems. In this thesis, we proposed models to develop solutions for the cold-start problem from two different perspectives. We aimed for a deterministic and a heuristic study that can be used in different scenarios. In the first perspective, we introduced a new heuristic framework that optimizes item-based similarity models to provide top-N recommendation lists using Continuous Ant Colony Optimization with a non-deterministic approach. Thanks to its heuristic structure, we aimed to create specific recommendation lists for users and change them according to each session, while at the same time aiming to balance the relevance of the user and the item variety in the recommendation lists. In the second perspective, we introduced two new Collaborative Filtering techniques deterministically. In the first model, we developed an asymmetric similarity matrix among the items based on the z-score normalization of the Gram-matrix we obtained using the implicit data, and in the second model, we aimed to reduce the sparsity with the item predictions with the assist our novel item similarity matrix, thus enabling more accurate decomposition of the latent factors in the user-item matrix we provided. We evaluated all of our methods on well-known datasets and observed that our methods outperform similar recommendation models in a variety of scenarios, including cold-start users, cold-start systems, and providing of unpopular product recommendations.
Predicting Covid-19 infection using machine learning and feature selection methods
The purpose of this thesis is to develop new COVID-19 prediction models using different machine learning and feature selection methods. Particularly, Multi-Layer Perceptron (MLP), Tree Boost (TB), Radial Basis Function Network (RBF), Support Vector Machine (SVM), and K-Means Clustering (KMC) have been used to construct various regular and feature selection-based COVID-19 prediction models. The minimum redundancy maximum relevance (mRMR) and Relief-F algorithms have been chosen as the feature selectors. The dataset has information related to 20.000 patients (i.e., 10.000 positives, 10.000 negatives) and includes several personal, symptomatic, and non-symptomatic variables. The accuracy, precision, recall, and F1-score metrics have been used to assess the models' performance, whereas the generalization errors of the models has been evaluated using 10-fold cross-validation. The results show that, in general, MLP outperforms all other ML classifiers for predicting the COVID-19 infection. The average performance of mRMR is slightly better than Relief-F in predicting the COVID-19 infection of a patient. The symptom-based variables such as fever, cough, and headache have been found as the most vital predictors of COVID-19 infection.
Data mining based on regularized convolutional neural network for time series: Financial prediction algorithm
This thesis aims to design a generalizable distance-based moving average (DBEMA) method for predicting time series. In our study, we focused on a specific area of financial time series. In order to increase the performance of prediction accuracy, DBEMA was combined with features selected by Recursive Feature Elimination (RFE) by using Classification and Regression Tree (CART) estimators and sequential feature selection (SFS) by using Gradient Boosting Machine (GBM). Although many artificial neural networks (ANNs) have been applied to a number of time series predictions and modelling, convolutional neural networks (CNN) have not been used much for time series prediction directly in literature and are still open to improvement. For predicting the trend of time series with DBEMA, time series are defined in the form of different time-lagged moving average patterns to identify the relations between each of them. The distances between moving averages (MA) and changes in their positions towards each other are examined for predicting future trends of time series. First of all, time series are defined so as to cover different time lags of 9 days, 50 days and 200 days in exponential moving average (EMA) forms and the distances between each of them and positions between each of them are marked. To improve the performance of the distance-based moving average method, CART and GBM algorithms are used for selecting better financial features in with RFE and SFS models, respectively. The combination of distance-based features and selected financial features are converted into 2-D images which are then classified by CNN. According to the experimental results, the proposed algorithm, CNN-DBEMA, outperforms other classification techniques in literature. Key Words: Distance-Based Features, Moving Average, Financial Time series Prediction, Convolutional Neural Network