Theses supervised by Prof. Dr. Selma Ayşe Özel
10 theses · Çukurova University
Effects of feature extraction techniques on classification of turkish texts
The purpose of this thesis is to determine the most effective method for extracting features and to develop an effective method of selecting features for the classification of Turkish documents in different types. We analyze the effects of preprocessing methods, weighting schemes, and feature selection on the performance of Turkish document classification. In the study, 5 different term weighting methods that are tf, tp, logtf, normtf, tf*idf are compared and it is found that "tf" and "tf*idf" give the best results. After that effects of stopwords removal are investigated, and it is observed that stopwords removal improves classification performance. Then we compare 5 different stemming algorithms that are Zemberek, Affix Stripping, Fixed Prefix 3, 5, and 7 to find out the effects of stemming algorithms on the classification. The results of classification obtained from applying stemming and using the raw form of terms are compared, and the raw form of terms gives more accurate classification results. The effects of n-gram based feature extraction, and feature selection methods that are our proposed standard deviation based method, well-known information gain, and chi-square algorithms are compared. The experimental results indicate that the n-gram feature extraction and standard deviation-based feature selection algorithms give the best results and these methods improve the classification accuracy positively.
Improving web page classification with unlabeled data
There are plenty of unlabeled data in different areas and effective ways are needed to be found to use them. In order to drive the useful information from these unlabeled data, semi-supervised learning methods are used. In this thesis, two different semi-supervised learning methods are proposed, namely Incremental Parallel Training with Cross-Validation (IPT-CV) and Incremental Serial Training (IST). The proposed semi supervised learning methods employ supervised classifiers and different views of the datasets for labeling unlabeled data efficiently. Therefore, to determine which classifiers and feature extraction algorithms should be used in the proposed semi-supervised learning methods experiments are performed. Then, to evaluate the effectiveness of the proposed methods, two known semi-supervised learning methods are implemented; Co-Training, and Iterative Cross-Training (ICT). Since web is a land of unlabeled files that is increasing tremendously, the web domain is chosen for the experiments. In the thesis, 13 binary classification datasets are used from the publicly available WebKB (i.e., Course, Faculty, Project, and Student), Banksearch (i.e., Biology, Commercial Banks, Motor Sport, and Programming), SyskillWebert (i.e., Bands, Biomedical, Goats, and Sheep) datasets, as well as manually collected Conference dataset. Experiments on 30 different randomly chosen initial labeled sets are made for each dataset and the results are analyzed statistically. According to these analyses, it is observed that the performance of the two proposed methods are very high, especially the IPT-CV method has the highest classifying performance among all methods.
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.
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.
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
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.
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
Detecting multilingual offensive language in social media using deep neural networks
The spread of offensive language on social media platforms has become an alarming reality in society today. The utilization of such language for the purpose of insulting and attacking people represents one of the most detrimental forms of online behavior. Its negative consequences extend to users across different communication platforms, significantly impacting their psychological and mental well-being. To combat this digital malady, data scientists and NLP researchers have taken the task of finding a solution. They have developed several classifier models employing machine learning and deep learning techniques, aimed at identifying several forms of offensive language within textual contexts. These models are designed to process the text by either removing offensive language or preventing its publication on the internet. This study seeks to address the issue by evaluating the performance of deep learning methods on a collected dataset that is formed by collecting a numerous amount of Arabic texts and labeling them. Additionally, comparison of the performance of different deep neural network classifiers namely, Convolutional Neural Network (CNN), Reccurrent Neural Network (RNN), and Long Short-Term Memory (LSTM), and a Language Model namely RoBERTa, performed on the Arabic dataset, as well as some additional datasets in English and Turkish languages, aiming to show the effects of different preprocessing on texts, feature selection and effectiveness of deep neural networks and Transformers across different linguistic texts. The results of this study suggest that RoBERTa is a strong candidate for various language, it achieved the highest validation accuracy across most datasets, showcasing its effectiveness for various languages and tasks. Additionally, an ensemble classifier combining RoBERTa and CNN is introduced and tested, demonstrating good results in improving classification performance.
Detecting offensive language from social media using word embedding and language models
This research addresses the pressing challenges posed by the proliferation of abusive content on social media platforms, tackling this issue in both English and Arabic languages. To construct a robust framework for detecting offensive language, we have employed cutting-edge methodologies. These include leveraging prominent language models such as Base BERT, Mini BERT, and GPT-2, as well as utilizing LSTM (Long Term Memory) models and an SVM (Support Vector Machine) classifier. Additionally, we have harnessed word embedding techniques like GloVe and Word2Vec to capture the intricate semantic relationships among words. The primary objective of this research is to fortify the detection mechanisms for offensive language, thereby nurturing a safer online environment, especially for vulnerable user groups like children and adolescents. Despite the relatively limited availability of Arabic resources for identifying offensive language, our research bridges this gap. It makes a substantial contribution to the field by presenting an extensive dataset encompassing various Arabic dialects. Through meticulous evaluation, we have optimized the synergy among the aforementioned methods to achieve precise classification of offensive content. In summary, this research aspires to cultivate a safer digital society and deepen our comprehension of the dynamics of offensive language within both Arabic and English social media spheres. Notably, in the English language, our best accuracy of 93.29% was achieved with the HateBERT model and Base BERT in tandem with an SVM classifier, employing a dropout rate of 0.4. For Arabic language, the highest accuracies were attained with the truncated dataset, achieving an accuracy of 89.35% when utilizing AraBERT Tweet, and with the entire dataset using AraBERT Tweet, reaching an accuracy of 92.86%. These achievements mark significant milestones in our pursuit of effective offensive language detection.
Identification of cyberbullying using machine learning techniques
The pervasive utilization of social media platforms has introduced a multitude of threats, among which cyberbullying stands as a significant concern. Cyberbullying is defined as the repetitive use of social and electronic media, to perpetrate immoral actions with the intent to inflict harm upon a victim. In both Turkey and globally, cyberbullying is recognized as a pressing issue necessitating urgent attention due to its profound emotional and psychological repercussions, which can include depression, stress, anxiety, and in extreme cases, suicide. For this concern, numerous researchers and scientists have endeavored to develop solutions leveraging machine learning (ML), deep learning (DL), and natural language processing (NLP) techniques. These efforts aim to create models capable of identifying cyberbullying within textual contexts, with the ultimate goal of either detecting and removing such content or preventing its dissemination. However, the large number of studies in this context have been conducted in English, with a dearth of research in Turkish language. Moreover, existing studies often segregate their analyses between English and Turkish, overlooking potential cross-linguistic nuances. Motivated by these gaps in the literature, this thesis endeavors to address the detection of cyberbullying in both English and Turkish languages. Framed as a binary text classification problem, the research scrutinizes the efficacy of standard techniques across languages, aiming to discern whether a unified approach can effectively detect cyberbullying in diverse linguistic contexts. To achieve this objective, comprehensive experiments are conducted, including the comparative evaluation of ML, DL, and Large Language model (LLM) Furthermore, an array of feature extraction techniques, including traditional feature weightings and word embeddings, are rigorously assessed. Additionally, the efficacy of an optimization technique LoRA, applied to LLM, is thoroughly evaluated. Furthermore, recognizing the problem created by labeled data shortage, a novel semi-supervised learning technique is proposed. Specifically, a self-training with LLM is implemented, demonstrating its potential to enhance classification performance by leveraging a blend of both labeled and unlabeled data, thereby mitigating the resource-intensive nature of manual labeling processes. This research contributes to advancing cyberbullying detection methods and encourages more inclusive approaches across languages to combat cyberbullying in the digital sphere.