Organizational Unit: Department of Data Science
Loading...
Date established
City
Country
ID
17 results
Publication Search Results
Now showing 1 - 10 of 17
Master ThesisPublication Metadata only Finding the determinants of player market value in association football using FIFA video game dataEren, Ozan Can; Özener, Okan Örsan; Özener, Okan Örsan; Özener, Başak Altan; Albey, Erinç; Yanıkoğlu, İhsan; Güler, M. G.; Department of Data Science; Eren, Ozan CanFutbol, son onyıllarda dünyadaki en popüler sporlardan biri haline gelmiştir. Artan popülerliği ile birlikte futbol, büyüyen milyar dolarlık bir pazar halini almıştır. Büyük Avrupa liglerindeki üst düzey oyuncuların transferleri için yüksek meblağlar ödenmektedir. Bu durum, futbolcuların piyasa değerlerine dikkat çekilmesine sebep olmuştur. Bu alandaki ucu açık konulardan birisi ise futbol oyuncularının piyasa değerlerini belirleyen faktörlerin tespitidir. Daha önce bu faktörlerin bir kısmı üzerine araştırmalar yapılmış olmasına rağmen odak noktası genellikle saha içi istatistikler olmuştur. Bu tez kapsamında benzersiz bir veri seti kullanılmakta ve futbol oyuncularının FIFA 2015, 2016, 2017 ve 2018 oyunlarındaki esas özellik ve niteliklerine odaklanılmaktadır. Oyuncu piyasa değerleri "transfermarkt.com" adresinden temin edildi. 2014-2015, 2015-2016, 2016-2017 ve 2017-2018 sezonlarının her biri için Avrupa'nın en büyük 5 liginde (İngiliz Premier Ligi, İspanyol La Liga, Alman Bundes Liga, İtalyan Serie A ve Fransız Lig 1) en az doksan dakika sahada kalan futbolcular analiz edildi. Çeşitli değişken seçme metotları kullanılıp Sıradan En Küçük Kareler modelleri oluşturularak defans, orta saha ve hücum bölgelerindeki spesifik alt pozisyonlarda oynayan oyuncuları değerli kılan en önemli demografik, takımsal, fiziksel, teknik ve mental özellik ve nitelikler her sezon için tespit edildi. Farklı alt pozisyonlarda (Stoper, Bek, Merkezi Defansif Orta Saha, Merkezi Orta Saha, Kanatçı ve Forvet) oynayan oyuncuları değerli kılan farklı özellikler ve nitelikler olduğu gösterildi. Her 4 sezon ve 6 alt pozisyon için oluşturulan Sıradan En Küçük Kareler modellerinin çıktıları paylaşıldı.Master ThesisPublication Metadata only Modern data management strategies for machine learning tasks: A sports analytics use case on cloud platformMete, Emrah; Albey, Erinç; Albey, Erinç; Özener, Okan Örsan; Güler, M. G.; Department of Data Science; Mete, EmrahThere is no doubt that data is the most valuable asset today. The efforts of enterprises in digital transformation and creating a data-driven culture are the most concrete indicators of this. Nowadays, where data transforms all industries, it is possible to follow the rapidly developing technological developments in this field. Appropriate data management strategies are the basis of creating data-driven organizations. When the evolution of data management architectures is examined, it is possible to say that the biggest factor that triggers this evolution is the changing and increasing data sources and the velocity of data production. In addition, with the increase in the importance of business use cases that need to be done in real time, it has become a very crucial need to process data quickly and turn it into action. Today when data is strategic importance, enterprises that can manage data correctly could gain competitive advantages. Being able to the correct data management can be built with the support of up-to-date and modern approaches. The infrastructures established by the integration of new and modern methods into the platforms turn into more agile structures. This increases the number of value added services to be produced from data by providing speed and flexibility to organizations. Today, the outputs expected to be produced from data management platforms go beyond descriptive and diagnostic analytic. Now, artificial intelligence, machine learning and data science are important parts of these platforms and these are opened new channels for the future of enterprises. In this thesis, basic needs and capabilities of modern data management architectures are described and detailed explanations were made on reference architectures in the industry. Besides, data management strategies and expectations were discussed. An example prototype of the data management platforms, which is explained in detail in this thesis, has also been developed on the cloud platforms. In this prototype, the entire life cycle of the data was considered and each step was developed in detail. In addition, a data science project was developed using the data collected on the platform. Thus, an end-to-end solution has been implemented.Master ThesisPublication Metadata only Machine learning framework to price-setting risk-averse data-driven newsvendor problemAtsız, Eren; Kayış, Enis; Kayış, Enis; Danış, Dilek Günneç; Albey, Erinç; Önal, Mehmet; Güler, M. G.; Department of Data ScienceMany business-owners, including manufacturers, retailers, distributors are in need to have decisions to make ordering for tons of products in their daily routines. This decision-making cannot be considered trivial since there is always a risk of ordering more or less then what is actually required, which may result in severe undesirable circumstances that are overstocking or shortages. The newsvendor problem in the literature focuses on this trade-off in its classical applications. In such approaches, there is an assumption of demand distribution is known. However, true demand is almost never known and constantly change due to being open to oodles of internal and external parameters; therefore, it is almost impossible to be known and hard to be predicted in real life. For this reason, traditional approaches to newsvendor model in the literature are very valuable yet it is possible to create different variations of this problem which are more appropriate for the real world instances. In addition to quantity of ordering, the determination of price is another serious concern of decision-makers. In order to satisfy the need of determining correct amount of order and regarding price positioning, there is a different approach to newsvendor problems, which is price-setting providing not only quantity but also price information. In this paper, we focus on this variant of newsvendor problem and create a data-driven price-setting newsvendor model, where demand is completely unknown. What is unique in this work is to integrate another element in addition to price decision, which is enabling a risk-averse decision-making perspective. In this way, the constructed model decides on quantity of order and price for products, as well as satisfies a certain risk constraint in order to a set a barrier of probability to not generate under a predetermined level of profit. The price-setting and risk-averse applications of newsvendor model exist in the literature yet combining these two different perspectives is what makes our approach novel.Master ThesisPublication Metadata only A revised approach to cryptocurrency portfolio optimization using advanced Q-learning and policy iteration frameworksAltok, Ceren; Albey, Erinç; Albey, Erinç; Önal, Mehmet; Güler, M. G.; Department of Data ScienceDespite all the factors that cause concern among investors, such as volatility and de centralization of crypto world, the popularity of cryptocurrencies continues to grow steadily. The cryptocurrency market still holds its allure for many investors due to the high profit levels it has experienced in the past. With the entrance of numerous alt coins into the market, portfolio management becomes much more challenging. In the literature, we come across numerous studies proposing efficient portfolio management techniques for cryptocurrencies. This study presents proposed models developed based on policy iteration and Q-learning algorithms. Under Q-learning, three distinct sub-models are introduced: Deep Q-Network (DQN), Double Deep Q-Network (DDQN), and Double Dueling Q Network (DDDQN). All of these models are trained using 6-month training periods and compared using 10 different training and testing periods. Additionally, to eval uate both of proposed policy iteration and Q-learning models, baseline models were created for each algorithm, and the performance of the proposed models was assessed against these baseline models. The results indicate that among Policy Iteration models, the proposed model has the highest average ROI value of 3%, making it the top-performing model. Similarly, among Q-learning models, the proposed DQN model surpasses both baseline models and other Q-learning models, with an average ROI value of 2%. Considering all the models, the proposed Policy Iteration model achieves the highest average ROI value, while the proposed DQN and the proposed DDDQN model demonstrates the lowest volatility in terms of ROI standard deviations.Master ThesisPublication Metadata only N-hop influence maximization problem under deterministic linear threshold modelOdabaşı, Sena; Danış, Dilek Günneç; Danış, Dilek Günneç; Kayış, Enis; Keskin, E.; Department of Data Science; Odabaşı, SenaThe Influence Maximization Problem (IMP) finds a set of highly influential nodes within a social network in order to maximize the spread of influence. We consider that people can have influence on their direct (1-hop), 2-hop and 3-hop neighbors. IMP with extended influence transitivity is called n-hop IMP. In this paper, we study the problem under the deterministic linear threshold model and propose a heuristic solution. In our proposed heuristic model, there are two parts, extended seed set algorithm and local search. Main purpose of extended seed set algorithm is creating a node set and send it to local search which is based on trying to improve it via replacing the nodes in the given set. We used two different features for selecting the candidate nodes. We propose an equation to estimate the value of a node set without actually computing. We used real-life and synthetic networks to test our solution method and generated weight and threshold values in three different methods.Master ThesisPublication Metadata only Optimizing inventory routing: an integrated machine learning solution approachAktaş, Taha Huzeyfe; Özener, Okan Örsan; Özener, Okan Örsan; Ekici, Ali; Yakıcı, E.; Department of Data ScienceInventory Routing Problem (IRP) arises from vendor-managed inventory business set tings where the supplier is responsible for replenishing the inventories of its customers over a planning horizon. In the IRP, the supplier makes the routing and inventory decisions together to improve the overall performance of the system. In our setting, the supplier’s goal is to minimize total transportation costs over a planning horizon while avoiding stock-outs at the customer locations. We assume that the supplier has a fleet of homogeneous capacitated delivery vehicles and abundant availability of the product to be delivered to the customers. Each customer has a constant de mand/consumption rate and limited storage capacity to keep inventory. To address this problem, we propose a novel integrated clustering and routing algorithm. In the clustering phase, we partition the customer set into clusters, ensuring that each cluster is served by a single vehicle. To accomplish this, we employ a novel deep learning model within the clustering framework. In the routing phase, we develop the delivery schedule for each cluster. What sets our approach apart is its consider ation of the three key decisions—when to deliver, how much to deliver, and how to route—by integrating both a mathematical model and a machine learning model in the decision-making process. We evaluate the performance of the proposed clustering and routing algorithms against existing literature, and our results demonstrate sig nificant improvements. Furthermore, the proposed neural network-based clustering approach serves as an effective representation of how machine learning algorithms can enhance decision-making structures.Master ThesisPublication Metadata only Machine learning based allocation in a lot sizing gameSarıoğlu, Ömer Berkay; Özener, Okan Örsan; Özener, Okan Örsan; Özener, Başak Altan; Çelikyurt, U.; Department of Data ScienceSupply chains often have confliction objectives and operate in a finite resource setting. Traditionally, companies focused on their internal processes to generate cost reduction opportunities in order to increase their profitability. However, recent studies suggest that collaboration is the key to have sustained benefits among supply chain partners. Supply chains, which usually have conflicting objectives, can form coalitions and take advantage of collective payoffs. In this paper, we analyze a collaborative production setting where several companies facing varying demands throughout a finite planning horizon attempt to reduce their procurement costs by ordering from a common supplier. The participants exploit the synergy among themselves by maximizing the capacity utilization of the supplier. We design a novel cost allocation method using various machine learning techniques with the goal of generating a cost-allocation mechanism that ensures the sustainability of the collaboration. We conduct a computational study to compare and contrast our proposed method with the generic methods in the literature. We discuss the advantages and disadvantages of these methods in terms of solution quality and computation time.Master ThesisPublication Metadata only Deep reinforcement learning approach for trading automation in the stock marketKabbani, Taylan; Duman, Ekrem; Duman, Ekrem; Albey, Erinç; Alkaya, A. F.; Department of Data Science; Kabbani, TaylanDeep Reinforcement Learning (DRL) algorithms can scale to previously intractable problems. The automation of profit generation in the stock market is possible using DRL, by combining the financial assets price "prediction" step and the "allocation" step of the portfolio in one unified process to produce fully autonomous system capable of interacting with its environment to make optimal decisions through trial and error. In this study, a continuous action space approach is adopted to give the trading agent the ability to gradually adjust the portfolio's positions with each time step (dynamically re-allocate investments), resulting in better agent-environment interaction and faster convergence of the learning process. In addition, the approach supports the managing of a portfolio with several assets instead of a single one. This work represents a novel DRL model to generate profitable trades in the stock market, effectively overcoming the limitations of supervised learning approaches. We formulate the trading problem as a Partially Observed Markov Decision Process (POMDP) model, considering the constraints imposed by the stock market, such as liquidity and transaction costs. More specifically, we design an environment that simulates the real-world trading process by augmenting the state representation with ten different technical indicators and sentiment analysis of news articles for each stock. We then solve the formulated POMDP problem using the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm and achieved a 2.68 Sharpe ratio on the test dataset. From the point of view of stock market forecasting and the intelligent decision-making mechanism, this study demonstrates the superiority of deep reinforcement learning in financial markets over other types of machine learning such as supervised learning and proves the credibility and advantages of strategic decision-making using DRL.Master ThesisPublication Metadata only Multilabel classification with neural networkEkşioğlu, Sezin; Özener, Okan Örsan; Özener, Okan Örsan; Özener, Başak Altan; Çelikyurt, U.; Department of Data Science; Ekşioğlu, SezinMulti-label classification has huge importance for several applications, it is also a challenging research topic. It is a kind of supervised learning that contains binary targets. The distance between multilabel and binary classification is having more than one class in multilabel classification problems. Features can belong to one class or many classes. There exists a wide range of applications for multi-label prediction such as image labeling, text categorization, gene functionality. Even though features are classified in many classes, they may not always be properly classified. There are many ensemble methods for classification. However, most of the researchers have been concerned about better multi-label methods. Especially little ones focus on both efficiency of classifiers and pairwise relationships at the same time to implement better multi-label classification. In this paper, we worked on modified ensemble methods by getting benefits from k-Nearest Neighbors and neural network structure sequentially to address issues beneficially and to get better impacts from the multi-label classification. Publicly available datasets (yeast, emotion, scene, and birds) are performed to demonstrate the developed algorithm efficiency, and the technique is measured. Our algorithm outperforms benchmarks for each dataset with different metrics. The result of the algorithm is competitive with the state-of-the-art results. Especially, in the weighted average of false-positive minimization and false-negative minimization, the algorithm passes the benchmarks.Master ThesisPublication Metadata only Forecasting the future price movement of cryptocurrency assets by convolutional neural networkÇetin, Aysel; Albey, Erinç; Albey, Erinç; Önal, Mehmet; Güler, M. G.; Department of Data ScienceDigital or virtual currency known as cryptocurrency uses cryptography for security and is not controlled by a central bank. Cryptocurrencies control the issue of new units and record transactions using decentralized technology, such as blockchain. Cryptocurrencies are entirely digital and have no physical form, in contrast to traditional currency, which is real and backed by a government or financial institution. Although Bitcoin was the first and best-known cryptocurrency, there are now thousands of other coins in use, including Ethereum, Tether, BNB, XRP etc. Bitcoin's value can be extremely unstable and it is frequently utilized as an investment or speculative asset. In some locations, it can also be used to make purchases of products and services, and some companies even accept it as payment. Due to the formation of this sector in relation to the growth in earnings and followers, all focus turned in the direction of the cryptocurrency market. Despite the abundance of studies that have been done in the past for this area, less image processing research has been done for the next movement prediction. This paper uses Bitcoin (BTC) dataset and tries to create a tool to project the upcoming price direction. The target variable is a binary type as the next movement will decrease or increase direction. The challenge of forecasting the next day's price movement involves learning from the information from the previous day by transforming them to the images. Since the main goal is using image processing for prediction, Convolutional Neural Networks, one of the most well-known deep learning techniques, and human judgment will be used. As a result, it is aimed to create an algorithm that makes a successful buy-sell decision on BTC and to achieve profitability. However, if the algorithm created for this study has adequate performance and structure on BTC, it will be simple to adapt it to other cryptocurrency kinds in the future.