Theses supervised by Doç. Dr. Sadettin Emre Alptekin
12 theses · Galatasaray University
Yapay sinir ağları ile yazılım projelerinin eforunun tahminlenmesi
The software industry is growing rapidly and gaining importance all over the world. Nearly all companies and institutions from various industries have software projects to develop new applications and platforms. As required with every project, accurate effort estimation has become a crucial problem for the companies, especially for project managers. Since 1970s different methods and models have been developed for estimating software projects' efforts. The first milestone model was COCOMO, which is a constructive method proposed in the late 1970s. Many different models followed, the most popular and usable models being Function Point and Use Case Point. After 2000s, due to advances in technology, Artificial Neural Networks has gained in importance especially among the problem domains that benefit from data analysis and self-learning. Software development effort estimation also share similar characteristics as there is typically old projects' data on hand that should help foresee new projects' efforts. Therefore, in this study we build a software estimation model by using neural network methodology. The features for the network were chosen as a result of an extensive survey. The applicability of the methodology is demonstrated via real-life software project data provided by one of the largest banks in Turkey. Keywords : Software development effort estimation, Neural networks, Back propagation algorithm
Otomatik insan ruh sağlığı asistanı: Pasif sensör verisinden stres tanıma çalışması
Stress level among people is rising through years and passive sensing data from mobile phones or other ubiquitous devices have started to found its place in applications of mental health observation. With the ultimate goal of creating an automatic human mental health assistant that helps people to have a better mental condition, a step is taken by creating a stress recognition model. In previous works, the researchers have found correlations between sensor data and mental health conditions and attempted to predict the stress level of the user. Due to there is no direct link between any sensor data with mental health, Machine Learning algorithms are employed to uncover relations with multiple sensors and mental well-being. The utilized machine learning algorithms for prediction work with non-sequence data hence the researchers need to extract features that represent historical sensor data with instant features. However, extracted features cannot completely represent a sequence of time data. Within the scope of this study, we showed that LSTM, CNN and CNN-LSTM algorithms which accept sequences of data as input and reaches exceptional performances in different applications can also work in passive mobile phone sensor data to predict human mental stress. The performance of the model on StudentLife dataset which includes passive mobile sensing data of college students has 62.83% accuracy on 460 test instances by training with 800 instances with LSTM model. Diversity and size of the data are very small and the data-hungry LSTM model could not generalize on adapted features with the small sample size. Although we did not adapt complex features, the results are promising and encourage us to improve data size and continue to research on this topic.
E-ticarete yönelik işbirliksel filtreleme tabanlı öneri sistemi
Recommender systems are one of the core engagement functions for e-commerce industry. In a typical recommender system, customer and product data is analysed and a prediction model is generated which evaluates products for prospective customers. In terms of business value, it helps individuals identify their interest among overwhelming variety of products. In this paper, a collaborative filtering based recommender system framework is proposed for Turkey's leading e-commerce platform hepsiburada. First of all, implicit feedback and customer-product prediction pairs are prepared from collected data. Second, a regularized singular value decomposition (SVD) based matrix factorization model is established for collaborative filtering (CF). Customers and products are represented with latent factor vectors. This model is trained with implicit feedback, as the SVD problem is solved with Alternating Least Squares (ALS). Third, predictions are gathered from CF model. Then, predictions are limited to ten-product recommendation sets. At last, recommendations are evaluated by behavioural data generated by prospective customers.
Türkiye çiçekçilik endüstrisi için sürdürülebilir kalkınma stratejilerinin ve sürdürülebilir finansman mekanizmalarının değerlendirilmesi
Sustainability concept, which is mostly associated with environmental approaches, has evolved into policies under the title of sustainable development throughout the world, and has been transformed into an action plan that will cover the present and future generations by addressing the economic and social dimensions. It is clear that agriculture plays an important role in sustainability. Floriculture, which is one of the significant agriculture-based industries, maintain its importance for many countries with its commercial position. Although Turkey has significant advantages, it can be seen that it's not a major player in the global floriculture market. In addition, there are serious public debates for quite some time, about the decline in the agriculture. The aim of this study is to determine strategies for the sustainable development of the Turkish floriculture industry, and thereupon to evaluate these strategies and sustainable financing mechanisms together with their economic, environmental and socio-political dimensions. Within the scope of the study, we have identified sustainable development strategies that can be applied to the industry, with the help of a comprehensive Strength-Weakness-Opportunities-Threats analysis, where we highlight the current situation of the Turkish floriculture industry, and individual expert opinions. The Analytical Network Process model that we developed in line with our goal, makes it possible to evaluate the sustainable floriculture approach under benefits, opportunities, costs and risks merits. We believe that the criterion system that we have identified and results we obtained, will draw an applicable strategic road map on behalf of Turkish floriculture and other similar industries.
Anormallik tespitinde çift yönlü üretken çekişmeli ağlar yaklaşımının geliştirilmesi için bir yöntem
Anomaly detection is considered as a challenging task due to its imbalanced and unlabelled nature. Numerous machine learning methods are applicable to the anomaly detection task. Conventional machine learning algorithms, such as supervised anomaly detection methods require labeled data sets and can obtain reasonable achievements on balanced data sets. However, they mostly suffer from the class imbalance problem. Unsupervised anomaly detection methods, on the other hand, assume that the more significant part of the data is normal and are inclined to label the least fit instances as anomalies. Semi-supervised methods create structures from normal data, which represents standard data distribution. To overcome this challenge, the combination of different machine learning approaches such as supervised, unsupervised, semi-supervised learning are proposed in the literature. With the advent of neural networks and generative models, different methodologies derived from neural networks are applied to anomaly detection tasks. In this study, we use the KDDCUP99 and Credit Card Fraud Detection data set, consider them as an anomaly detection task, and implement Bidirectional Generative Adversarial Networks, considering it as a one-class anomaly detection algorithm. Since generator and discriminator are highly dependent on each other in the training phase, to reduce this dependency, in this paper, we propose three different training approaches for Bidirectional Generative Adversarial Networks by adding extra training steps to it. We also demonstrate that proposed approaches increased the performance of Bidirectional Generative Adversarial Networks on anomaly detection task.}
Tedarik zincirlerinde süreç iyileştirmek için blokzinciriteknolojisinin bulanık QFD yöntemiyle değerlendirilmesi
With the increasing complexity in globalized business environments, businesses encounter a significant challenge to optimize and innovate their processes to take advantage of a highly competitive market. With its unique characteristics, blockchain technology will change the way we handle business processes. This study aims to support the blockchain technology selection decision of a supply chain considering the requirements of its stakeholders, primarily focusing on its procurement process. We made a case study application seeing the problematic supply chain metrics and software-related characteristics of the case supply chain using an integrated Cognitive Map-QFD-TOPSIS methodology to select an appropriate enterprise blockchain alternative. Vague or imprecise process and metric definitions and imprecise quality definitions can create problems in the technology selection process, so these definitions should be set in a standard form. We used Supply Chain Operations Reference (SCOR) modelling and Business Process Modelling and Notation (BPMN) language to deal with the vague and imprecise process and metric definitions. Furthermore, trade-offs are the most crucial decisions that must be made in an architectural design. To prioritize SCOR metrics and focus on specific metrics without losing information, we used Fuzzy Cognitive Maps since there are trade-offs between SCOR performance metrics. For global standardization, the International Organization for Standardization (ISO) and the International Electro-technical Commission (IEC) established information technology guidelines. In this study, ISO/IEC 25010-25012 standard software quality attributes were evaluated for focusing on problematic software attributes of the case supply chain. To prioritize ISO software characteristics and blockchain characteristics separately, we again applied Fuzzy Cognitive Maps. It is highly significant to determine complex relationships between ISO software 2 characteristics considering their contribution to the software quality. However, software stakeholders in the case supply chain cannot easily interpret the changes that the complex relationships between software characteristics can create. To the best of our knowledge, in literature, there is not a study evaluating the relationships between the software characteristics from both ISO/IEC 25010 and ISO/IEC 25012 standards using Fuzzy Cognitive Map. After inspecting all blockchain characteristics from the literature, we summarized blockchain characteristics to be able to support blockchain selection decisions of a selected enterprise. We also evaluated trade-offs between blockchain characteristics using Fuzzy Cognitive Maps. In literature, blockchain technology studies in supply chains are either too supply chain-specific or too technical in software topics. However, supply chain and software related topics need to be systematically linked to each other for the selection of a suitable blockchain alternative. In the end, we selected an appropriate enterprise blockchain alternative using Fuzzy (Technique for Order Preference by Similarity to Ideal Solution) TOPSIS. There is no QFD-Cognitive MapTOPSIS integrated study on enterprise blockchain selection, specifically inspecting the trade-offs between blockchain characteristics. In this study, SCOR metrics, ISO characteristics and metrics, blockchain design requirements, and the selection of alternative blockchains enabled us to support to-be processes of a selected supply chain. Blockchain became more understandable by using a standardized language in quality characteristics and performance dimensions (SCOR, ISO, etc.), also limited in the literature. Additionally, blockchain characteristics were evaluated in detail to be able to support enterprise blockchain selection decisions. Fuzzy Cognitive Map-based scenario analysis enabled us to consider the different effects of different scenarios in our study. Keywords: Procurement process, Blockchain technology, Fuzzy Cognitive Map, Fuzzy QFD, Fuzzy TOPSIS
Değişken otokodlayıcı kullanarak envanter stoğundaki anomalilerin tespiti
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 companies try to detect and prevent inventory record inaccuracy caused by employee or customer theft, damage or spoilage and wrong shipments. Our study aimed to detect inaccurate stocks using the Variational autoencoder (VAE) method, and we used the real inventory stock data of one of Turkey's largest supermarket chains. This method learns the distribution of data, and it is a great advantage to use this in data that changes over time. The VAE learns the usual pattern from normal time series data and detects anomalies by identifying the unseen data pattern, possibly reducing time and effort while gathering error data. In addition, this method can be applied to any product level. However, we use the interquartile range method to define the threshold for our model; therefore, it becomes parametric. 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. We show how to handle real-life data noises, missing values etc. The experimental findings show that the proposed approach can detect anomalies in the low and high inventory stock quantity and quickly apply to other time series anomaly detection problems.
Psikoloji bilimi yaklaşımıyla öznel iyi hal durumunun makine öğrenmesiyle modellenmesi
Recent advances in pervasive computing enable the collection of personal health-related data using diverse sensors in the everyday-life environment. However, human behavior modeling and analysis, particularly the quantification of subjective well-being, is still challenging, as there are variations in its definition and measurement. The psychology literature defines different perspectives on subjective well-being, such as hedonic and eudaimonic. In this thesis, we propose a model for predicting an individual's subjective well-being from the psychological perspective using her/his daily activities collected via smart wristbands, social relationships monitored through smartphones, and personality traits data from surveys. The model is applied to the NetHealth study, a heterogeneous data set of 577 student participants from the University of Notre Dame. We developed a multi-class classifier based on commonly accepted machine learning algorithms. The results enable us to predict an individual's well-being with almost 80% accuracy. We show the feasibility of a pervasive application as a personalized well-being assistant.
Karakter özellikleri ve sensör tabanlı günlük aktivite verisi kullanılarak öznel iyi oluş tahmini
People enjoy different things in life. For example, while some people enjoy walking and socializing, others may get the same pleasure from sleeping. Enjoyment improves people's moods, and different metrics are used in psychology to measure that well-being level. One of the most widely used is the "subjective well-being" scale. The degree of this scale is measured by short questionnaires applied to individuals. In this study, subjective well-being was chosen as the target. We transformed the target to a binary scale (0-1), and the problem is considered as classification. In the model created, the individuals' subjective well-being was estimated by applying Decision Tree, Logistic Regression, Naïve Bayes, SVM, KNN and Ensemble classification methods. The person's daily physical activity, socialization, and big five personality traits were used as attributes. As a result of the models applied, we have achieved an accuracy between 64 and 80%. Unlike other studies, this study revealed a model that predicts 80% accuracy by composing sensor data with character traits. In this study, NetHealth dataset that collected from college students during consecutive periods was used.
Güç transformatörleri üretiminde adam-saat tahmini için Gaussian proses regresyon, destek vektör makineleri ve ANFIS modellerinin karşılaştırılması
Production times affect the product valuations because it is directly relevant to the production cost. Therefore, accurate estimation of the production times is one of the key problems to be able to make correct product valuations and pricing. The man-hour unit is widely used for this purpose and it is taken into consideration while making product valuations before tendering phase of projects. It is especially important in tailor-made production which generally has a labor-intensive manufacturing environment because each product is a different project. When the man-hours are predicted by experts instead of systematic tools derived from local data, they often result in incorrect predictions which affect the ex-factory cost of the product. If the incorrect prediction has a negative deviation from the actual man-hour, it may result in a reduction of the profit margin which is a serious problem for all kinds of businesses or if the man-hour prediction is over-calculated to prevent this problem then another problem emerges and the cost-effectiveness may be lost for customers in tendering. Hence, the prediction of the man-hour without excess under-estimations and over-estimations based on systematic methods in labor-intensive manufacturing environments is important to be able to keep the overall profitability of factories. There have been several kinds of research to overcome the problems resulting from incorrect expert estimations in different industries such as shipbuilding and the aircraft industry. Also, it has been researched under the title of "effort estimation" in the software development field. However, there has not been any study directed to the Power Transformer manufacturing which has a tailor-made production mentality. In this study, man-hour predictions in Power Transformers manufacturing have been studied based on data-driven methods and machine learning applications. The results showed that the proposed forecasting system can be a good alternative to the existing ones.
Öznel iyi oluş modelini kullanarak stress ölçeğine dayalı akıl sağlığı tahmini
It has been decided that one of the most significant facets of existence is well-being, as this is the greatest way to keep people healthy and productive. This has led to the conclusion that well-being is one of the most important parts of existence. Because there are a substantial number of different ways to define "well-being," there are also a significant number of different scales that can be used to assess it. The purpose of the well-being scale is to provide general information on the level of an individual's quality of life. In this study, Subjective Well Being (SWB) characteristics are used in conjunction with machine learning classifiers to make predictions regarding levels of stress. In order to collect data from college students between the years of 2015 and 2019, sensors and self-reports were utilized, and approximately 700 students took part. To have a glance of understanding human nature and the requirements, it has has recently emerged as a popular topic of study among experts. Because of this, the combination of Artificial Intelligence (AI) and Psychology can appear to be straightforward; yet, in order to identify each of the aspects and their reflections in AI's research, the subject of psychology needs to be thoroughly researched. The purpose of this study is to provide individuals with a basic framework for understanding their physical and mental health situations by presenting and explaining in detail the findings of a psychological research. A growing number of studies have been carried out to investigate the possibility of accurately predicting a person's level of happiness by making use of carefully constructed models. In order to build a Subjective Well-Being (SWB) model that has any chance of success, it is necessary to conduct research into the histories of the features. We have selected the variables from the literature on SWB that are appropriate for the real-world data instructions, and these variables come from the suitable categories. The purpose of this work is to evaluate the model by providing it with SWB characteristics and then classifying the different levels of stress using machine learning methods in order to determine how well the model functions when applied to a real dataset. We have achieved significant metric scores, which may be taken into account for a particular task, despite the fact that it is a multiclass classification issue.
Toplu sipariş problemi çözümünde değişken talep için hesap tablosu yöntemi
A proper replenishment strategy is a critical enabler in the success of increased revenue, net profits and customer service. Inventory management requires constant and careful evaluation of external and internal factors and control through planning and review. In order to control the inventory, one major requirement is to provide efficient replenishment technique such as jointly replenishment of products. Searching for efficient replenishment techniques is a common and usually a mandatory topic for the organizations. We consider inventory systems with multiple products in the presence of volatile demand and jointly incurred order setup cost. In this thesis, a new adaptation of spreadsheet heuristic for volatile environment is presented. The simplicity of application of spreadsheet method and its efficiency enables us to consider its modified version for the joint replenishment problem under volatile demand. The principle of the procedure is to find a balance between the replenishment and holding cost for jointly replenished items. The real business data is conducted to evaluate the performance of the heuristic. The study shows that the proposed algorithm performs well in comparison with well known RAND heuristic for the numerical data. Additionally, the importance of using volatile demand strategy over deterministic strategy is also highlighted with a calculation. The proposed strategy gives higher customer service level which means lower unmet customer demand. Owing to the simplicity and effectiveness of the proposed algorithm, we believe that it can be applicable in volatile demand environments. Keywords: Inventory management, Spreadsheet heuristic for volatile demand, Joint replenishment problem.