Yüksek Lisans Tezleri
Permanent URI for this collectionhttps://hdl.handle.net/20.500.11779/1785
Browse
91 results
Search Results
Master Term Project Segmentation With Unsupervised Learning: an Application Using the Walker's Data(MEF Üniversitesi Fen Bilimleri Enstitüsü, 2021) Polat, Taylan; Özgür ÖzlükIn this project, the Walkers suitable for the service were filtered by using the dataset shared by the DogGo company. Then, unsupervised machine learning methods such as K-Means, Gaussian, Principal Component Analysis were used to score and cluster the most suitable walkers according to performance, willingness, and experience.DogGo is the first mobile application in Turkey that provides pet walking and grooming services to its customers in a safe and professional manner. DogGo provides a professional service where dogs are taken care of in dog families' own homes or at the caretaker's home for any need of dog families. DogGo Company wants to provide the best matching of walkers and animals, using Machine Learning algorithms, through a 5-step acquisition process for their walkers.While the results of the K-means models created on the unique sliders were compared with the help of the Elbow method and the Silhouette score, the results of the Gaussian models were compared with the AIC and BIC method. In addition, an RFM scoring in a classical structure has also been created. When the results of the study were examined considering the Elbow and Silhouette scores, it was shown that the model created with K-Means gave the best results, and the number of clusters was decided as 2.Master Term Project Game Recommendation System for Steam Platform(MEF Üniversitesi Fen Bilimleri Enstitüsü, 2021) Bayram, Serhan; Semra AğralıIncreasing number of choices and competition in the markets, force companies to differ in services they provide to their customers. Offering better services have a positive impact on customer loyalty, and to do so, companies should understand their customers’ interests and act accordingly. One popular method for this purpose is building recommendation engines to make personalized suggestions. In this project, collaborative filtering methods with implicit feedback are used to make recommendations to users of theSteam platform. The recommendation systems are built using two different matrix factorization techniques, Alternating Least Squares and Bayesian Personalized Ranking. Different models are created with implicit playtime data of the users and the results are evaluated by using Precision at k metric. Additionally, similar items that are offered by the models are analyzed. Results show that the models are considerably successful at finding personal choices and similar items. The best model finds the item in the libraries of 33% ofthe users.Master Term Project Online Shopping Purchasing Prediction(MEF Üniversitesi Fen Bilimleri Enstitüsü, 2021) Kazezyılmaz, İdil; Evren GüneyThis project aims to understand the purchasing behavior of the consumers and make predictions about purchasing according to website metrics such as page values, bounce rates.An existing dataset is used in this project. This dataset is available in the collection of data from an e-commerce website by Google Analytics, which consists of 10 numerical and 8 categorical attributes coming from 12,330 sessions. The 'Revenue' attribute is used as the class label. The attributes that have high impact on the prediction are; "Administrative", "Administrative Duration", "Informational", "Informational Duration", "Product Related" and "Product-Related Duration". They represent the number of different types of pages visited by the visitor in that session and the total time spent in each of these page categories.The "Bounce Rate", "Exit Rate" and "Page Value" features represent the metrics measured by Google Analytics for each page in the e-commerce site. The "Special Day '' feature indicates the closeness of the site visiting time to a specific special day (e.g. Mother’s Day, Valentine's Day) in which the sessions are more likely to be finalized with a transaction.Since the purpose of this project is to predict potential purchasing using existing data, in the prediction part several machine learning algorithms such as decision trees, random forests will be applied to compare the models. The most suitable model will be chosen among these algorithms.Master Term Project Rfm Based Customer Segmentation for a Mobile Application(MEF Üniversitesi Fen Bilimleri Enstitüsü, 2021) Baykan, Ozan Barış; Özgür ÖzlükIn this project, customer segmentation was made for Doggo, a mobile application that brings together trained dog walkers for people who are not able to provide daily needs of their dogs. The data was organized by obtaining the columns of recency, frequency, monetary and tenure, and RFM-based customer segmentation was made using machine learning algorithms such as K-means and Gaussian Mixture Model (GMM). Then, the model was built with the part of the dataset that includes recency, monetary and tenure columns using K-means. In addition, with a function developed, the RFM and tenure will be repeated at intervals determined by the Doggo operation team, and this tool is used to monitor the customer condition changing. Various marketing campaigns have been proposed according to the current situation and the transitions they have made.Master Term Project Employee Performance Prediction(MEF Üniversitesi Fen Bilimleri Enstitüsü, 2021) Sivas, Barış; Özgür ÖzlükDogGo is a company that aims to provide safe and professional dog walking and grooming services to dog owners through the mobile application. Thanks to the DogGo application, dog owners and people who is employee of company and wants to walk their dogs (to be called Walkers) can meet on the same platform on the mobile application interface. The problem was determined by company that they needed to be able to accurately predict the performance of the walkers in the upcoming dog-walker matches, thus ensuring the correct dog walker match. This study will be planned to serve to this company for calculating their current walkers’ performance in an accurate way. The relevant machine-learning model will first be based on the manual scoring system made by the company for the performance of existing employees, and then the model will be developed in the light of the gains obtained from this. For the performance of the model, the employees and their characteristics are important for the first time.Master Term Project QPICAR Deep Learning(MEF Üniversitesi Fen Bilimleri Enstitüsü, 2021) Beğde, Özge; Tuna ÇakarThe aim of the project is to train a smart tool kit named "Sunfounder Raspberry Pi Robot Car" to move without hitting the walls in a closed area. The goal is to maximize the driving time without crashing by reducing the number of hits. Ultrasonic sensor data collected from the vehicle are processed with reinforcement learning and deep reinforcement learning algorithms and the results are compared. In this study, Python programminglanguage is used. In this study, firstly, the Q-Learning method, which is a reinforcement learning algorithm based on Markov decision processes, is used. The method basically relies on a memory table, Q-Table, in which the Q-values of the agent moving from one state to another are kept. This table is updated according to the results of the Bellman equation in every action of the agent, and as a result of this iterative process, it is optimized to provide that the agent moves to maximize its rewards. Deep Q-Learning (DQN) is used as a deep reinforcement learning algorithm. This algorithm was developed by the DeepMind Technologies team in 2013. Basically, it is based on the use of the Bellman equation, which is an element of the Q-Learning method, incombination with neural networks. This method is often used for training agents in complex and multidimensional environments such as video games. Due to the different type of the data used on the algorithm, minor changes were made to adapt it to the study. RElu and Softplus are used as activation functions. The results of the training process show that the DQN algorithm has an important advantage in terms of training the agent in a short time. At this point, the results are in accordance with other academic studies demonstrating the success of the DQN algorithm for complex environments.For future work, by differentiating the equipment that collects data on the vehicle, different data types such as image, temperature value, oxygen value can be collected and processed. At the same time, with changes to the reward setup in the algorithm, the agent can be trained to move to a specific target or to take actions to avoid a specific target.Master Term Project Ad Click Prediction Using Machine Learning Algorithms(MEF Üniversitesi Fen Bilimleri Enstitüsü, 2021) Uncu, Nazlı Tuğçe; Hande KüçükaydınOnline advertising has a great potential to boost business’ revenue. One of the key metrics that defines the success of online ad campaigns is click through rate (CTR) which indicates the total number of clicks received in relation to the total impression. Therefore, the click prediction systems, which have the aim of increasing the click through rates of online advertising campaigns by predicting the clicks, have become essential for businesses. For this reason, predicting whether an advertisement will receive a click fromthe user or not attracts the attention of researchers from the both industry and academia. In this capstone project, the click prediction is studied by using Avazu’s click logs dataset. The effects of having high cardinality categorical features and imbalanced data are examined during data preprocessing phase and then relevant features are selected to be used in modeling. The methods that are used for this classification problem are decision trees, random forest, k-nearest neighbor, extreme gradient boosting, and logistic regression. According to the results of the study, extreme gradient boosting shows the best performance.Master Term Project Credit Card Froud Detection Using Machine Learning(MEF Üniversitesi Fen Bilimleri Enstitüsü, 2021) Erdoğan, Tibet; Duygu Taş KütenThis project aims to find the most efficient machine learning models to detect fraudulent transactions on credit cards. The dataset used for this project consists of credit card transactions made by European cardholders in September 2013. This dataset presents transactions that have occurred in two days, where there are 492 frauds out of 284,807 transactions. Machine learning methods, such as decision trees, logistic regression and random forest classifier are used to predict the fraudulent transactions. Performance of these machine learning models are compared to achieve the highest accuracy. According to the results, it is found that the random forest classifier is the most effective model, and the SMOTE technique used to overcome the data imbalance performs better than the under-sampling technique. It is also observed that the models employed with the under-sampled data misclassify large number of non-fraud transactions as fraud. Lastly, by means of the random forest with the over-sampling technique (SMOTE), it is observed that the feature “V13” has the most important role in detecting fraud.Master Term Project The Effect of Bert-Based Grammatical Analysis on Google Search Results(MEF Üniversitesi Fen Bilimleri Enstitüsü, 2021) Çolak, Oğuz; Özgür ÖzlükThis study aims to study the BERT, namely Bidirectional Encoder Representations from Transformers model, which is introduced by Google and is of great importance in content analysis, and to examine the role of grammatical accuracy in the process of content quality measurement and Search Engine Results Pages (SERP). BERT has an important role among the algorithms used by Google in order to maintain the quality of search results and to provide more relevant content to users by understanding the content more effectively.In this study, CoLA data, which is accepted as the most reliable data in this field and therefore used frequently in similar BERT studies, is used. The main purpose here is to make a BERT-based grammatical evaluation of sentences in a content and then examine these results on pages with optimal ranking values, to examine the connection between search results and grammatical accuracy and the importance of this parameter.In this context, the project consists of two phases. In the first phase, the content of the pages that are visible in the first 20 in 50 different queries are scored with the pre-trained BERT model. In the second phase, a dataset that includes different SEO-focused metrics of the same pages is created manually, and the importance of the BERT score among these features is investigated.Master Term Project Predicting the Price of Bitcoin: Using Machine Learning Time Series Methods(MEF Üniversitesi Fen Bilimleri Enstitüsü, 2020) Ulutaş, Sezer; Utku KoçCryptocurrencies have greatly increased their Bitcoin-led popularity in recent years due to increased trading volumes and massive capitalization in the market. These cryptographic forms of money are not just utilized for exchanging nowadays, they are additionally acknowledged for fiscal exchanges. It appears to be evident that financial specialists, dealers and people, in general, are progressively intrigued by bitcoin and altcoins as costs rise and the arrival on ventures made increments. This examination centres around applying estimate models that will make precise value forecasts forcryptographic forms of money. The data were taken from two different exchanges and evaluated as combined dataset. As a result of the evaluation, it was determined that the prices were close to each other in terms of value and the data were combined. We obtained the daily time series data by determining the Bitcoin weighted price as a dependent variable and Open, Close, High, Low and Volume as independent variable. We predicted the next 6 months with ARIMA, LSTM and XGBoost methods. We compared these estimates using MSE, MAE, MAPE and R squared performance metrics. LSTM is the model with the best R squared value of 29.7%. In the process performed by taking the average of LSTM, XGBoost and ARIMA performed with the name of Average ML method, the R square value was found to be 41.6% as a much better result than LSTM.