Theses supervised by Dr. Öğr. Üyesi Tuna Çakar
11 theses · MEF University
Customer churn prediction for the Pay-TV sector
Understanding the reasons for customer churn provides added value in terms of retaining existing customers, as customer attrition leads to revenue loss for companies and incurs marketing costs for acquiring new customers. In this study, the 6-month historical data of a Pay-TV company operating in Turkey was used, and due to the imbalanced nature of the dataset on a label basis, the oversampling method was applied. During the model development phase, various artificial learning algorithms (Random Forest, Logistic Regression, K-Nearest Neighbors, Decision Tree, AdaBoost, XGBoost, Extra Tree Classifier) were utilized, and their performances were compared. Based on the evaluation of success criteria for each model, it was observed that the tree-based Random Forest, Extra Tree Classifier and XGBoost achieved the highest performance for this dataset.
Fraud detection and prediction with machine learning applications
The main purpose of this study is to determine the fraudulent activities on transactions of the customers of a company that is active in the factoring sector, and accordingly, to capture measurable parameters with exploratory data analysis based on the historical transaction and connection data of the customers, and then to perform predictive models for the target. A hit rate of around 79% was achieved in XGBoost and CATBoost models, which are classification model algorithms. In this way, it is aimed to directly detect fraudulent activities on a trasnaction basis by acting in a more effective, efficient and correct approach after detecting the customer that shows high potential to make fraud.
Serverless vs. on-premises: A performance analysis of ml deployment with aws fargate, GCP Cloud run, and On-Prem
In this study, I present a comparative analysis of the changes occurring during the deployment process of machine learning models, both in On-Premises systems and cloud service providers. The successful deployment of machine learning models holds critical importance for businesses and organizations aiming to enhance their productivity. Understanding and comparing how models behave in different environments is of paramount significance to make informed decisions. Prominent commercial organizations like AWS and GCP offer reliable and cost-effective cloud services tailored to provide customized web applications. Our primary objective in this article is to guide cloud customers by highlighting the key features of the most recognized Cloud Service Providers and facilitating informed decision-making through comparisons with the On-Premises option. Additionally, I explore the advantages of managed services such as AWS Fargate and Google Cloud Run, which streamline application deployment. Through this research, my goal is to offer useful insights that help companies succeed in the fast-paced, cutthroat business environment by helping them make wise strategic decisions.
EAFT: Evolutionary algorithms for GCC flag tuning
The runtime of written codes is a matter of great importance, especially for code that is compiled once and executed multiple times. It is very important for developers to ensure that the resources required by a code are used as efficiently as possible, and that the runtime is as low as possible. Developers who use compilers such as GCC or LLVM to compile and run code written in C or C++ can optimize their code manually and, with certain optimization pointers, are able to make it run faster. This will provide the shorter runtime, but completıng this manual optimization is within the abilities of every developer since determining the right combination from more than 200 flags requires significant expertise. Many studies have tackled this issue. In this study, Evolutionary Algorithms for GCC Flag Tuning (EAFT) have been developed as a solution to this problem. This Autotuner, which is completely open-source, runs the code provided by the end user according to the specifications also selected by the end user, and searches for the most suitable optimization markers. For the code to be given In line with this study, which specifically addresses the end user, the user can input the code path directly from the Terminal, as well as specify the selection method and the crossover to be used. These choices can be made without the need to alter the code. The genetic algorithm and particle swarm optimization to be used is also presented to the user in EAFT, and unlike in other studies, genetic algorithm contain not one but several models.
Customer transaction predictive modeling via machine learning algorithms
Bu çalışmanın temel amacı faktoring sektöründe aktif olarak faaliyet gösteren bir şirketin müşterilerinin davranışlarını ve özelliklerini saptayabilmek, buna bağlı olarak da müşterilerin geçmiş verilerinden yola çıkarak, keşifçi veri analiziyle ölçülebilir parametreler yakalayabilmek ve akabinde hedefe yönelik tahminsel modellemeler gerçekleştirebilmektir. Sınıflandırma modeli algoritmalarından SVM ve Extra Trees modellerinde %80 seviyesi üzerinde isabet oranı yakalanmıştır. Bu sayede yüksek potansiyel gösteren, yani sorgulattığı çeki, işletme olarak onayladıktan ve teklif yapıldıktan sonra kabul etme ihtimali yüksek olan müşterileri tahmin edilmesi daha etkin, verimli ve doğru yaklaşımlar içerisinde hareket edip aksiyon alarak, işletme bazında işlem hacmine doğrudan katkısı sağlanması amaçlanmıştır.
Predicting the preference of liking using fNIRS and machine learning algorithms
The fMRI method, which is generally used to detect behavioral patterns, draws attention with its expensive and impractical features. On the other hand, the near-infrared spectroscopy (fNIRS) method is less expensive and portable, but it is as effective as fMRI in creating a good prediction model. With this method, a model has been developed that can predict whether a person likes a visual stimulus or not, using various classical machine learning algorithms including Support Vector Machines (SVM), Random Forests, XGBoost, LightGBM and K-Nearest Neighbors (KNN). With implementing tree-based and booster algorithms in addition to SVM and KNN which have been frequently used algorithms in this fNIRS domain, it was aimed to do a complementary comparison in addition to these acknowledged algorithms. Moreover, various missing value imputation methodologies were used to find the best suitable approach for this kind of classification problem. K-Means clustering, which is an unsupervised learning method, was also utilized to cluster similar fNIRS measurements of participants that may improve classification results by one-hot encoding those groups. Furthermore, certain feature extraction and wrapper methodologies were also applied for an attempt to enhance the performance of liking prediction models as a secondary goal. PCA, Isomap and t-SNE methodologies were implemented as feature extraction approaches, and forward selection wrapper design was utilized as an additional step to further development of the model by comparing their scores with each other. Cross-validation F1-scores of these models were used to find out the best model among them. Leave-one-group-out cross validation was exploited in comparison of the models. This meant that these cross-validation results corresponded to each of participants' data i.e. testing every participants' fNIRS measurements alone in each fold. This way both every score specific to each participant could be seen and it ensured models' results were statistically reliable. Following evaluations also included permutation and Wilcoxon Signed-Rank tests to compare each model's performance with each other by testing the statistical significance of those results. Keywords: machine learning, decision-making, optical brain imaging, fnirs, feature extraction, feature selection
Comparing audio features for speech emotion recognition using machine learning algorithms
Voice is an integral part of our lives. The demand for voice technology in both art and human-machine interaction systems has recently been increased. More information can be transferred quickly by voice. Speech is a natural way of communicating and as a result of this, it is primarily preferred for contacting users in technological areas. Our voice conveys both linguistic and paralinguistic messages in the course of speaking. The paralinguistic part, for example, rhythm and pitch, provides emotional cues to the speaker. Emotions consist of cognitive, physiological and behavioural changes and all these phenomena are interrelated. Generally, an emotion is a state that affects the thoughts and is capable of determining behaviour. Emotion also creates physical and psychological changes. Speech Emotion Recognition topic examines the question 'How is it said?' and an algorithm detects the emotional state of the speaker from an audio record. Within the scope of this study, machine learning models are developed with classification methods to resolve the problem of speech emotion recognition. Voice consists of a lot of characteristics. However, the optimal audio feature set related to the emotional state cannot be determined yet. The main aim in this study is obtaining the most distinctive emotional features. For this purpose, in order to compare audio features based on different domains Root Mean Square Energy (RMSE), Zero Crossing Rate (ZCR), Chroma and Mel Frequency Cepstral Coefficients (MFCC) features are examined for emotion recognition. A pre-trained model namely wav2vec Large which has been developed more recently is used to create the inputs also. Support Vector Machine, Multi-Layer Perceptron and Convolutional Neural Network techniques are utilized for developing learning models for comparing traditional features and the pre-trained model representations. In this paper emotions namely, Happy, Calm, Angry, Boredom, Disgust, Fear, Neutral, Sad and Surprise are classified, and furthermore, the models are trained and tested with English and German speech datasets. When the classification results are examined, it is concluded that the most successful predictions are obtained with the pre-trained representations. The weighted accuracy ratio is 91% for both Convolutional Neural Network and Multilayer Perceptrons models while this ratio is 87% for the Support Vector Machine models. Among the emotional states, Fear has the highest recognition ratio with 95% f-score with Convolutional Neural Network technique which uses a pre-trained model.
Customer-merchant shopping behavior modeling in e-commerce
The rapid advancement of e-commerce has necessitated the adoption of sophisticated analytical models to decipher customer behaviors and evaluate seller performance. In this study, we propose a novel 'Customer-Seller Shopping Behavior Model' that integrates customer preferences, purchasing patterns, and seller product offerings. The model utilizes customer metrics such as RFM (Recency, Frequency, Monetary), category preferences, and customer lifetime value, alongside seller data on category focus and sales performance, to facilitate segmentation and personalized recommendations. Employing K-Means, hierarchical clustering, and decision tree algorithms, customers and sellers are categorized based on their behavioral similarities. Furthermore, predictive models and hybrid recommendation systems are implemented to forecast customer purchasing propensities and match them with suitable sellers. The results demonstrate that the proposed model significantly enhances customer engagement, seller effectiveness, and overall platform efficiency. This research contributes to a data-driven, scalable, and dynamic analytical framework for e-commerce platforms.
Comparing the effectiveness of graph neural networks and machine learning algorithms for fNIRS-based neuromarketing research
Functional near-infrared spectroscopy (fNIRS) has some advantages over other brain imaging methods in terms of cost and portability. For this reason, its use in neuromarketing is increasing. However, fNIRS brings some challenges along with its advantages. Due to features such as multichannel measurement and high temporal resolution, the nature of fNIRS data is complex and multidimensional [7]. Neuromarketing researchers have utilized machine learning algorithms to overcome these challenges. When these studies are analyzed, it is seen that successful results have emerged. Machine learning has influenced researchers working on graphs as well as neuromarketing researchers. Thus, graph neural networks have emerged, which allow the application of artificial neural networks to graph data structures. Thanks to the fact that the brain can be modeled as a graph structure using functional connections [14] and the high temporal resolution of fNIRS [7], there are neuroimaging studies using graph neural networks and fNIRS together. However, despite successful results, there is no neuromarketing research using this combination. Therefore, in this study, the performance of graph neural networks in fNIRS-based neuromarketing was analyzed and compared with machine learning algorithms that have been shown to yield successful results in this context. For the comparison, fNIRS measurements of a neuromarketing experiment conducted to determine perceptions toward brands were used. In the experiment, consumers were asked to decide (yes/no) whether the adjective shown with the brand logo was appropriate for the brand. The data set was obtained by cleaning the obtained measurements. First, a supervised machine learning approach was applied to this dataset. After the dataset went through several data preprocessing stages, various algorithms were trained on it. These were K-Nearest Neighbors, Support Vector Machines, Random Forest, Naive Bayes, and XGBoost algorithms. Then, two different voting classifiers, one for soft voting and one for hard voting, were created from the algorithms that were more successful than the others. After the machine learning approach was completed, the graph neural network approach was applied. The data obtained through fNIRS was transformed into a graph structure using functional connections in the brain. The Pearson correlation coefficient was used to calculate the functional connections. Since a graph was created for each trial of the participants and each graph had a label (yes/no), classification was performed at the graph level. For graph classification, the generated graphs were given as input to graph neural network architectures. The architectures used in the study consisted of Graph Convolutional Network, Graph Attention Network, and Graph Isomorphism Network. Finally, a soft voting classifier was created by combining these architectures. Test accuracy values of all methods were calculated and binomial confidence intervals were added to these values. The comparison results showed that machine learning algorithms generally outperform graph neural networks. Additionally, machine learning models based on ensemble learning have the best scores.
Customer churn prediction for a personal care product retail chain operating in Turkey
Understanding the reasons for customer loss and the customer behaviors leading to it, as well as being able to predict customer's loyalty to an industry or a company provides enormous advantages in retaining existing customers and avoiding revenue loss due to the marketing and advertising costs associated with attracting new customers. In this study, the 29-month data from a personal care product retail chain operating in Turkey was used, and because of the imbalanced values and non-customer entries of the dataset, the oversampling method and synthetic sampling was applied. During the model development phase, Logistic Regression, Decision Tree, K-Nearest Neighbors, Random Forest, Extra Trees Classifier, and MLP (Multi-Layer Perceptron) Classifier were applied, and their performances were evaluated using metrics such as accuracy, recall, F1-score, precision, and confusion matrix. Based on these comparisons, it was observed that the Random Forest and MLP Classifier models demonstrated the best performances for this dataset, while other tree-based algorithms, such as the Extra Trees Classifier and Decision Tree, achieved slightly lower but comparable performance.
Forecasting for e-commerce sales using supervised machine learning algorithms
The burgeoning landscape of e-commerce relies significantly on predictive analytics to drive operational efficiency and strategic decision-making. This thesis delves into the theoretical underpinnings of machine learning algorithms, showcasing their evolution and pivotal role in facilitating the growth of online commerce. At its core, this research centers on forecasting sales patterns through the analysis of an extensive e- commerce dataset. Forecasting stands as a linchpin for various critical functions within e-commerce enterprises. Its multifaceted applications encompass inventory management, ensuring optimal stock levels and streamlined deliveries, financial planning through astute asset management, dynamic pricing strategies, and the enhancement of customer satisfaction via efficient delivery operations. Furthermore, forecasting plays a pivotal role in refining marketing endeavors, enabling tailored campaigns and judicious budget allocation. The integration of machine learning algorithms fortifies these functionalities. Central to this research is the foundational task of sales prediction in the e- commerce realm, with a specific emphasis on integrating campaign variables. Leveraging six diverse machine learning algorithms, the study aims to discern the most accurate and explicable model. Remarkably, the investigation identifies LGBM as the most suitable algorithm. Notably, the inclusion of campaign variables, an aspect seldom explored in prior studies concerning forecasting, yields intriguing insights. However, contrary to initial presumptions, the SHAP analysis reveals a lesser influence of campaign variables on the model's interpretability. Acknowledging this limitation, the study highlights the potential for augmenting model interpretability by employing clustering algorithms to effectively represent variables, as outlined in the limitations section.