Software Engineering Department

In order to effectively manage and retrieve the information comprised in vast amount of text documents, powerful text mining tools and techniques are essential. In this paper we evaluate and compare two state-of-the-art data mining tools for clustering high-dimensional text data, Cluto and Gmeans. Several experiments were conducted on three benchmark datasets, and results are analysed in terms of clustering quality, memory and CPU time consumption. We empirically show that Gmeans offers high scalability by sacrificing clustering quality while Cluto presents better clustering quality at the expense of memory and CPU time.

Text documents are usually unstructured and written in natural language. To apply conventional data mining techniques on text documents, a preprocessing operation is indispensable. In this paper, we introduce PRETO, a cross-platform, powerful and scalable preprocessing tool developed specifically for preprocessing Turkish texts, with a wide range of preprocessing options like stemming, stopword filtering, statistical term filtering, and n-gram generation. We demonstrate the performance and
scalability of PRETO with some experiments on large document collections.

In order to apply data mining techniques on texts that are written in natural language, a preprocessing step is used to transform this unstructured data into structured format. Stemming is an important text preprocessing technique. In this study, we empirically examined the impact of essentially three stemming methods on clustering Turkish texts: Zemberek, Affix Stripping, and Fixed Prefix. Although very similar results were obtained with all these methods in terms of clustering quality, Zemberek and Fixed Prefix 5 methods produced better results when compared to the others. Besides clustering quality, Zemberek and Fixed Prefix 5 methods are preferable stemming methods for Turkish text clustering applications due to high dimensionality reduction rate they provide.

In this study, we investigated 6,307 master’s and PhD theses prepared in the field of computer science and engineering between 1983 and 2012 in Turkey. Thesis data was compiled from the YÖK national thesis database web portal. We presented the profile analysis of the theses in terms of several attributes like thesis year, university, and advisor. Furthermore, we analyzed the popularity of research areas over the years, and presented the change in the research themes in Turkey based on the keywords obtained from graduate theses.

The advances in digital signal processing technology has led the use of speech processing in many different application areas like speech compression, enhancement, synthesis, and recognition. In this thesis, the issue of speech recognition was studied and a speaker dependent, large vocabulary, isolated word speech recognition system was developed for Turkish Language.

A combination of two common approaches to speech recognition problem was used in the project: Acoustic-phonetic approach and stochastic approach. The phonemes modeled by two-state Hidden Markov Models (HMM) were used as the smallest unit for recognition. Mel-Frequency Cepstral Coefficients (MFCC) was preferred as the feature vector extraction
method. A new algorithm was devised for phoneme detection and segmentation used in the training stage. Using phoneme-based recognition, the words that are not trained can be recognized by the system.

Keywords: Turkish speech recognition, phoneme based speech recognition, Hidden Markov Model (HMM), Mel-Frequency Cepstral Coefficients (MFCC), speech feature vector.

In this study, we examined 6,834 master's and PhD theses conducted on computer science and engineering between 1994 and 2013 in Turkey. Thesis data were compiled from the YÖK national thesis database web portal. We used text mining techniques to extract research concepts and their co-occurrence data from graduate thesis abstracts. Then, we applied social network analysis techniques on the concept co-occurrence networks to visually explore core research concepts, and connections and relationships among them. We showed that text mining and social network analysis techniques together were very effective for knowledge discovery in scientific documents in computer science and engineering domain.

Thanks to advances in information and communication technologies, there is a prominent increase in the amount of information produced specifically in the form of text documents. In order to, effectively deal with this “information explosion” problem and utilize the huge amount of text databases, efficient and scalable tools and techniques are indispensable. In this study, text clustering which is one of the most important techniques of text mining that aims at extracting useful information by processing data in textual form is addressed. An improved variant of Spherical K-means algorithm named multi-cluster spherical K-means is developed for clustering high dimensional document collections with high performance and efficiency. Experiments were performed on several document data sets and it is shown that the new algorithm provides significant increase in clustering quality without causing considerable difference in CPU time usage when compared to Spherical K-means algorithm.

Veritabanlarında Bilgi Keşfi olarak da adlandırılan Veri Madenciliği, veritabanları ve veri ambarları gibi çeşitli veri depolarında saklanmakta olan büyük miktardaki verinin işlenerek içindeki geçerli, daha önceden bilinmeyen, potansiyel olarak kullanışlı, yararlı ve değerli olabilecek bilginin çıkartılması sürecidir. Veri madenciliğinin önemli tekniklerinden biri olan veri kümeleme, benzer veri nesnelerinin farklı gruplara sınıflandırılması işlemidir. Bir veri kümesi belirli bir uzaklık ölçütüne göre alt kümelere ayrılırken hedef her bir altkümedeki nesnelerin ortak bir karaktere sahip olmasıdır. Metin Kümeleme ise doküman koleksiyonlarının doküman benzerliklerine bağlı olarak kümelere ayrıştırılmasıdır. Sonuç olarak, bir küme içerisindeki dokümanların genellikle benzer bir konuda olmaları beklenir.

Bu tez çalışmasında, metin veri setlerinin kümelenmesi için geliştirilmiş algoritma ve yaklaşımlar ayrıntılı olarak incelenmiş, çok boyutlu ve çok büyük metin veri setlerini kümelenmesinde karşılaşılan sorunlar ve zorluklar irdelenerek bunlar için çözüm yöntemleri getirilmiştir.

Tez çalışması dört bölümden oluşmaktadır. İlk bölümde öncelikle veri madenciliği, metin madenciliği ve metin kümeleme kavramları tanımlanarak metin kümelemedeki güncel problemlerden bahsedilmiştir. Daha sonra, tez kapsamında çözüm getirilen problemin tanımı yapılarak tezin amacı ve organizasyonu verilmiştir.

Problemin tanımının ardından ikinci bölümde metin madenciliği sistemlerinin genel yapısı, metin kümeleme, metin kümelemede karşılaşılan sorunlar, önişleme teknikleri, doküman temsil yöntemleri, boyut indirgeme teknikleri incelenmiştir. Ayrıca metin kümeleme sürecinin kalitesini ve başarımını doğrudan etkileyen uzaklık ve benzerlik ölçüm yöntemleri ile kümeleme kalitesi ölçüm yöntemleri ele alınmıştır. Bu bölümde son olarak metin kümeleme alanında yapılmış önemli çalışmaların, önerdikleri yaklaşım ve çözüm yöntemleriyle kapsamlı olarak incelendiği geniş bir literatür taramasına yer verilmiştir.

Tez kapsamında, çok boyutlu doküman koleksiyonlarının yüksek başarımla ve verimli bir şekilde kümelenebilmesi amacıyla, keskin kümeleme yapan Küresel K-Means algoritmasında değişiklik yapılarak esnek kümeleme yaklaşımlarındaki örtüşen kümeler oluşabilmesi fikri Küresel K-Means algoritmasına uygulanmış, K-Means döngüsü içerisinde dokümanların kümelere benzerliklerine göre belli ölçüde birden çok kümeye dahil olmasına izin verildiği özgün bir algoritma olan Çoklu-Küme Küresel K-Means algoritması geliştirilmiştir. Üçüncü bölümde, geliştirilen bu yeni kümeleme algoritmasının yapısı ayrıntılı olarak açıklanmış ve analiz edilmiştir. Çeşitli metin veri setleri üzerinde gerçekleştirilen deneylerle Çoklu-Küme Küresel K-Means algoritmasının işlemci zamanı kullanımında hissedilir bir fark oluşturmadan kümeleme kalitesinde büyük bir artış sağladığı, temel alınan Küresel K-Means algoritmasının ölçeklenebilirliğini koruyarak büyük doküman koleksiyonları üzerinde uygulanabilir olduğu gösterilmiştir.

Dördüncü bölümde, tez çalışması ile geliştirilen algoritmanın genel değerlendirmesi yapılarak sağlanan bilimsel katkılar özetlenmiş, konu ile ilgili çalışacak araştırmacılar için öneriler verilmiştir.

Tablet computers are frequently used in the education of children with autism. Studies show that educational activities involving tablet computers have positive effects on the acquisition of target skills by children with autism. In this research, the effectiveness of the use of tablet computers in teaching children with autism to use schedules was evaluated. Three male students of ages 5 to 7 diagnosed with autism participated in the research. A single-case multiple-probe-design-across-participants was used in the research. Independent variable of the research was the education provided via tablet computers using the graduated guidance technique, and the dependent variable was the acquisition of schedule-following skill. Follow-up and generalization data were collected on the 7th, 14th, and 21st days after the teaching sessions ended. Follow-up sessions were completed with 100% performance for all three subjects. Generalized correct responses were found 90% for one subject, and 100% for the other two. The results indicated that the subjects acquired, maintained, and generalized the schedule-following skills after the teaching sessions completed. In addition, the teachers of the subjects were asked their opinions about the application. These opinions that form the social validity data of the research support the research findings.

It is possible to extract valuable insights about the functional properties of a system by identifying and inspecting the community structure in the network that models the system. Community detection aims to extract these community structures from networks. Many community detection methods have been proposed that consider the problem from different perspectives. However, with the emergence of very large and complex networks from variety of domains, there has been a growing need for community detection methods that can operate at scale effectively and efficiently. Considering this, we propose a novel algorithm for large-scale community detection, based on two novel similarity indices we propose as well. In the first stage of our proposed algorithm, we generate candidate communities using a mechanism similar to information propagation very rapidly. Then, we merge small candidates that have fewer nodes than a calculated threshold with the larger ones using similarity between nodes and communities. Next, we engage a refinement operation on the candidates by moving all nodes to the candidates to which they are most similar using the same similarity index again. After that, we merge small communities with larger ones by using the similarity between communities until no gain in the modularity is obtained. Finally, in the last stage, we employ the same refinement operation as in the third stage. With an extensive experimentation on real-world and artificially-generated benchmark networks, we demonstrate and verify the performance and effectiveness of the proposed algorithm comparing it with the state-of-the-art methods. Experimental results indicate that our algorithm scales very well with growing size and complexity of networks. Besides, our algorithm outperforms most state-of-the-art community detection methods both in detection performance and computation time.

Delays in flights and other airline operations have significant consequences in quality of service, operational costs, and customer satisfaction. Therefore, it is important to predict the occurrence of delays and take necessary actions accordingly. In this study, we addressed the flight delay prediction problem from a supervised machine learning perspective. Using a realworld airline operations dataset provided by a leading airline company, we identified optimum dataset features for optimum prediction accuracy. In addition, we trained and tested 11 machine learning models on the datasets that we created from the original dataset via feature selection and transformation. CART and KNN showed consistently good performance in almost all cases achieving 0.816 and 0.807 F-Scores respectively. Similarly, GBM, XGB, and LGBM showed very good performance in most of the cases, achieving F-Scores around 0.810.

Thanks to advances in information and communication technologies, there is a prominent increase in the amount of information produced specifically in the form of text documents. In order to, effectively deal with this " information explosion " problem and utilize the huge amount of text databases, efficient and scalable tools and techniques are indispensable. In this study, text clustering which is one of the most important techniques of text mining that aims at extracting useful information by processing data in textual form is addressed. An improved variant of Spherical K-means algorithm named multi-cluster spherical K-means is developed for clustering high dimensional document collections with high performance and efficiency. Experiments were performed on several document data sets and it is shown that the new algorithm provides significant increase in clustering quality without causing considerable difference in CPU time usage when compared to Spherical K-means algorithm.

It is possible to extract valuable insights about the functional properties of a system by identifying and inspecting the community structure in the network that models the system. Community detection aims to extract these community structures from networks. Many community detection methods have been proposed that consider the problem from different perspectives. However, with the emergence of very large and complex networks from variety of domains, there has been a growing need for community detection methods that can operate at scale effectively and efficiently. Considering this, we propose a novel algorithm for large-scale community detection, based on two novel similarity indices we propose as well. In the first stage of our proposed algorithm, we generate candidate communities using a mechanism similar to information propagation very rapidly. Then, we merge small candidates that have fewer nodes than a calculated threshold with the larger ones using similarity between nodes and communities. Next, we engage a refinement operation on the candidates by moving all nodes to the candidates to which they are most similar using the same similarity index again. After that, we merge small communities with larger ones by using the similarity between communities until no gain in the modularity is obtained. Finally, in the last stage, we employ the same refinement operation as in the third stage. With an extensive experimentation on real-world and artificially-generated benchmark networks, we demonstrate and verify the performance and effectiveness of the proposed algorithm comparing it with the state-of-the-art methods. Experimental results indicate that our algorithm scales very well with growing size and complexity of networks. Besides, our algorithm outperforms most state-of-the-art community detection methods both in detection performance and computation time.

In this study, we analyzed undergraduate program preferences of students by using complex network analysis techniques. We collected program preferences data from the YokAtlas portal provided by the Council of Higher Education using a web crawler we developed. We constructed a kind of co-occurrence network we called co-preference network of 622 nodes and 6,136 edges from the collected raw data. We performed a comprehensive exploratory complex network analysis on the co-preference network using Cytoscape and NodeXL tools. Using several node centrality measures, we identified the most popular programs that students frequently preferred together with other programs. In addition, we observed the clusters of programs embedded in the network using several network community detection methods. Finally, we performed a structure analysis to compare our network to a corresponding random network, and we showed that our network had the common characteristic properties that many real-world network...

In recent years, there has been an increasing interest in research and application of agile software development. In academic literature, the number of systematic literature reviews (SLRs), systematic mapping studies, and unsystematic reviews has also increased. One of the aims of this article is to create a tertiary study of SLRs on agile software development research topics between 2013 and 2018. Our second goal is to provide an in-depth analysis with data including reference, author, institution, etc. belonging to this tertiary study, and to provide an in-depth analysis for software engineering researchers and practitioners to reveal the relationships of developments in this field. We applied a two-stage research method within the scope of the study. First, we obtained and examined the systematic literature reviews published in the field of agile software development from Web of Science, Science Direct, Scopus, and IEEE academic databases. Second, we visualized the publications e...