1,721,017 research outputs found
WikiHow veri seti üzerinde cümle gömmeleri ile soyutlamalı özetleme.
No full textSummarization is a well known natural language processing task that is used in our day-to-day lives. The field saw recent research using neural networks and word embeddings. We use WikiHow dataset and show that we can match performance of a similar model using sentence embeddings, and using abstractive summarization. We show that we can use sentence embeddings and lower input data size without impacting performance too much.Thesis (M.S.) -- Graduate School of Natural and Applied Sciences. Computer Engineering
Bluetooth düşük enerji teknolojisi (BDE) kullanarak yüksek doğruluklu yer tesbiti.
No full textUsing wireless technologies, locating objects with very precise measures is hard. It is even harder when we want to locate indoor positions due to surrounding materials preventing signals to be received. With the help of Bluetooth low energy technology, we can locate indoor objects/people with more precision than using previous Bluetooth technologies. In this thesis, methods will be improved and implemented to achieve cost-effective precise indoor location.Thesis (M.S.) -- Graduate School of Natural and Applied Sciences. Computer Engineering
E-ticaret verisi üzerinde gizli anlamsal analiz ve veri madenciliği yöntemleri kullanılarak öneri sistemi geliştirilmesi.
No full textRecommender systems are developed to provide better recommendations to users of e-commerce applications. In addition to this goal, e-commerce applications benefit from their recommender systems to show advertisements to users, apply discounts on specific items. The task of recommending an item to a user is always a challenge; luckily, there are many methods developed to complete this task such as collaborative filtering, association rule mining etc. These methods mainly look at the co-occurrence of items; however, we think that user behavior on different items should be extracted by doing latent semantic analysis on the data. Latent semantic analysis is used for understanding the context of a text, we think that it can be used for providing recommendations by processing transactional data. The data used throughout this thesis work consists of transactions made in various e-commerce companies. In this thesis work, existing methods and proposed recommendation methods are examined and recommendation results on this data are shown.Thesis (M.S.) -- Graduate School of Natural and Applied Sciences. Computer Engineering
Yapay zeka destekli 5G ve ötesi mobil şebekeler: güvenilirlik, hareketlilik içgörüleri ve sürdürülebilir işletim
No full textThe advancement of 5G and the emerging vision of 6G are transforming mobile networks into intelligent and adaptive infrastructures, presenting significant opportunities alongside considerable technical and operational challenges. In 6G networks, ubiquitous integration of artificial intelligence (AI) is expected across all layers, enabling autonomous decision-making, proactive management, and unprecedented service personalization. This thesis explores how AI can address key challenges in three vital domains of mobile network operations. It investigates AI-driven approaches for predicting radio link failures to enable proactive interventions, minimizing disruptions and ensuring seamless connectivity even under adverse conditions. It further examines how AI can analyze and model human mobility patterns from network data, generating critical insights that support disaster response, urban planning, and public safety, particularly in the wake of large-scale emergencies. Lastly, it focuses on AI-based strategies to optimize energy consumption in mobile networks, developing intelligent methods to reduce power usage without compromising coverage or service quality. Collectively, these contributions underscore AI’s transformative role in shaping the future of mobile communications, positioning next-generation networks to meet escalating demands for performance, resilience, and sustainability, while contributing to societal welfare and environmental responsibility.5G teknolojisindeki ilerlemeler ve 6G vizyonunun şekillenmesi, mobil şebekeleri akıllı ve uyarlanabilir altyapılara dönüştürmektedir. Bu süreç, yeni fırsatlar sunarken aynı zamanda teknik ve operasyonel zorlukları da beraberinde getirmektedir. 6G şebekelerinde, yapay zekanın (YZ) tüm katmanlara yaygın bir şekilde entegre edilmesi beklenmekte olup, bu sayede otonom karar alma, proaktif yönetim ve benzeri görülmemiş düzeyde hizmet kişiselleştirmesi mümkün olacaktır. Bu tez, YZ’nin mobil şebeke operasyonlarının üç kritik alanındaki zorluklara nasıl çözüm getirebileceğini incelemektedir. Çalışmada, olumsuz koşullar altında dahi kesintisiz bağlantı sağlamak ve aksaklıkları en aza indirmek amacıyla, radyo bağlantı hatalarını öngörmeye yönelik YZ tabanlı yaklaşımlar araştırılmaktadır. Ayrıca, YZ’nin şebeke verilerinden insan hareketliliği kalıplarını modelleyerek, özellikle büyük ölçekli acil durumların ardından afet yönetimi, kentsel planlama ve kamu güvenliğine katkı sağlayacak kritik içgörüler üretebileceği ortaya konulmaktadır. Son olarak, mobil şebekelerde enerji tüketimini optimize etmeye yönelik YZ tabanlı stratejilere odaklanılmakta; kapsama alanı veya hizmet kalitesinden ödün vermeden enerji kullanımını azaltacak akıllı yöntemler geliştirilmektedir. Bu katkılar bir arada değerlendirildiğinde, YZ’nin mobil iletişimin geleceğini şekillendirmedeki dönüştürücü rolü vurgulanmaktadır. Yeni nesil şebekelerin, artan performans, dayanıklılık ve sürdürülebilirlik taleplerini karşılayacak şekilde konumlandırılmasının yanı sıra, toplumsal refah ve çevresel sorumluluğa da katkı sağlaması hedeflenmektedir.Ph.D. - Doctoral Progra
Makine öğrenmesi metotları kullanılarak örtülü kanalların tespiti.
No full textA covert channel is a communication method that misuses legitimate resources to bypass intrusion detection systems. They can be used to do illegal work like leaking classified (or sensitive) data or sending commands to malware bots. Network timing channels are a type of these channels that use inter-arrival times between network packets to encode the data to be sent. Although these types of channels are hard to detect, they are not used frequently due to their low capacity and sensitivity to the network conditions. However, upcoming technologies like 5G and WiFi 6 offer more reliable networks with low latency, which we believe can work in favor of network timing channels and attract hackers to them. Therefore, we also believe that the detection of network timing channels is an increasingly important issue. In this thesis, we worked with two types of network covert channels: Fixed Interval and Jitterbug. Fixed Interval defines an inter-arrival time for each symbol to be transmitted and send network packets accordingly. On the other hand, Jitterbug does not create new packet traffic; it just delays existing packets for some predefined time. Two channels are very different: Jitterbug creates traffic that is similar to the legitiv mate network though has lower capacity, and Fixed Interval has a very different traffic shape from the legitimate network but has higher capacity. Our work has shown it is indeed possible to detect these channels with a decision tree with four features called mean, variance, skewness and kurtosis. However, more research is needed to make this system work in the real world.Thesis (M.S.) -- Graduate School of Natural and Applied Sciences. Computer Engineering
Going Beyond Counting First Authors in Author Co-citation Analysis
Full text linkThe present study examines one of the fundamental aspects of author co-citation analysis (ACA) - the way co-citation
counts are defined. Co-citation counting provides the data on which all subsequent statistical analyses and mappings
are based, and we compare ACA results based on two different types of co-citation counting - the traditional type that
only counts the first one among a cited work's authors on the one hand and a non-traditional type that takes into
account the first 5 authors of a cited work on the other hand. Results indicate that the picture produced through this non-traditional author co-citation counting contains more coherent author groups and is therefore considerably clearer. However, this picture represents fewer specialties in the research field being studied than that produced through the traditional first-author co-citation counting when the same number of top-ranked authors is selected and analyzed. Reasons for these effects are discussed
KATEGORI ̇LENDI ̇RME AMAÇLI KAPALI KUTU MODELLERDEN AÇIKLANABI ̇LI ̇R KURAL SETLERI ̇ ÇIKARIMI
No full textIn this work, we have proposed a new method and ready to use workflow to extract simplified rule sets for a given Machine Learning (ML) model trained on a classifi- cation task. Those rules are both human readable and in the form of software code pieces thanks to the syntax of Python programming language. We have inspired from the power of Shapley Values as our source of truth to select most prominent features for our rule sets. The aim of this work to select the key interval points in given data in order to extract if-then rule sets representing the black box models. We are able to generate rules that can mimic four different ML models on three datasets and one Deep Learning model on stock price dataset. We have evaluated promising similarity scores (around 90%) between the extracted rules and the ML models.Bu çalıs ̧mamızda, kategorilendirme üzerine çalıs ̧an kapalı-kutu modeller için açıkla- nabilirlik metodu gelis ̧tirdik ve sizlere sunuyoruz. Modeli açıklamaya yönelik üretmis ̧ oldug ̆umuz kural setleri hem insan hem de makina tarafından okunabilir formattadır. Gelis ̧tirdig ̆imiz metod içerisinde Shapley deg ̆erlerinden faydalandık ve bu deg ̆erle- rin önemli gördüg ̆ü aralıkları kendi kural setimizi çıkarmakta yardımcı oldular. Dört farklı makine ög ̆renmesi metodunu üç farklı veri seti üzerinde ve bir derin ög ̆renme metodunu finansal veri seti üzerinde test ettig ̆imizde, ümit verici sonuçlar elde ettik. Çıkarım yaptıg ̆ımız kural setleri ile orijinal model arasinda %90 civarlarında benzer- lik gözlemledik.M.S. - Master of Scienc
Derin Öğrenme Kullanarak Göğüs Hastalıkları Sınıflandırması İçin Görsel Açıklamaların Kararlılık ve Kalite Değerlendirmesini Geliştirme
No full textDeep learning models are renowned but their complex internal workings often render them opaque. While efforts are underway to amplify their explainability, a substantial gap persists. One pivotal challenge is the unavailability of quantifiable metrics for evaluating visual explanations, leading to a reliance on manual assessments or suboptimal metrics. This restricts scalability, compromises reproducibility, and undermines trustworthiness. The necessity for objective metrics is further underscored during hyperparameters fine-tuning and the evaluation of various Explainable Artificial Intelligence (XAI) models. Another prominent hurdle is the instability exhibited by models like Local Interpretable Model-agnostic Explanations (LIME). The random perturbations intrinsic to LIME lead to inconsistent explanations, eroding trust and obstructing their integration into critical applications.
To navigate these challenges, we unveil a robust method for the objective assessment, refinement, and juxtaposition of visual explanation algorithms, offering a concrete solution to metric inadequacy. Addressing instability, we introduce MindfulLIME. This cutting-edge approach leverages graph-based pruning and uncertainty sampling, strategically crafting purposeful samples and elevating the reliability and consistency of visual explanations - a discernible advancement over LIME. Our intricate qualitative analysis, applied to unseen random samples from the VinDr-CXR dataset, attests to the preeminence of our selected metric. In a comparative analysis involving multi-label, multi-class diagnosis of thorax diseases, MindfulLIME, evaluated via this refined metric, epitomizes optimal stability and augmented localization accuracy. This underscores its capacity to elevate the trust quotient associated with machine learning applications in the nuanced field of medical imagery.Derin öğrenme modellerinin karmaşık iç işleyişi onları anlaşılmaz hale getirmektedir. Açıklanabilirliklerini artırmaya yönelik çabalardaki önemli zorluklardan biri, görsel açıklamaları değerlendirmek için ölçülebilir ölçümlerin mevcut olmamasıdır. Bu durum, manuel değerlendirmelere veya optimal olmayan ölçümlere güvenilmesine yol açmaktadır. Bu, ölçeklenebilirliği kısıtlar, tekrarlanabilirliği tehlikeye atar ve güvenilirliği zayıflatır. Objektif ölçümlerin gerekliliği, hiperparametrelerin ince ayarı ve çeşitli Açıklanabilir Yapay Zeka (XAI) modellerinin değerlendirilmesi sırasında daha da vurgulanır. Öne çıkan bir diğer engel ise Yerel Yorumlanabilir Model-Agnostik Açıklamalar (LIME) gibi yöntemlerin sergilediği istikrarsızlıktır. LIME yöntemine özgü rastgele karışıklıklar tutarsız açıklamalara yol açarak güvenin aşınmasına ve kritik uygulamalarda kullanılamamasına neden olur.
Bu zorlukların üstesinden gelmek için, görsel açıklama algoritmalarının nesnel değerlendirmesi, iyileştirilmesi ve yan yana getirilmesi için metrik yetersizliğine somut bir çözüm sunan sağlam bir yöntem ortaya koyuyoruz. İstikrarsızlığa çözüm bulmak için MindfulLIME yöntemini tanıtıyoruz. Bu yeni yaklaşım, grafik tabanlı budama ve belirsizlik örneklemesinden yararlanarak, stratejik olarak amaca yönelik örnekler oluşturur ve görsel açıklamaların güvenilirliğini ve tutarlılığını artırır. Bu, LIME yöntemine göre gözle görülür bir ilerlemedir. VinDr-CXR veri setindeki görünmeyen rastgele örneklere uygulanan nitel analizimiz, seçtiğimiz metriğin üstünlüğünü kanıtlamaktadır. Toraks hastalıklarının çok etiketli, çok sınıflı teşhisini içeren karşılaştırmalı bir analizde, önerilen hassas ölçüm metriği aracılığıyla değerlendirilen MindfulLIME, optimal stabilite ve artırılmış lokalizasyon doğruluğu ortaya koymaktadır. Bu durum, tıbbi görüntülerle makine öğrenimi uygulamalarında güven oranını artırma kapasitesinin altını çizmektedir.Ph.D. - Doctoral Progra
Çevreci yeni nesil ağlarda pekiştirmeli öğrenme kullanarak güç optimizasyonu yapılması.
No full textThe next generation mobile networks have to provide high data rates, extremely low latency, and support high connection density. To meet these requirements, the number of base stations will have to increase and this increase will lead to an energy consumption issue. Therefore ``green'' approaches to the network operation will gain importance. Reducing the energy consumption of base stations is essential for going green and also it helps service providers to reduce operational expenses. However, achieving energy savings without degrading the quality of service is a huge challenge. In order to address this issue, we propose a machine learning based intelligent solution that also incorporates a network simulator. We develop a reinforcement based learning model by using deep deterministic policy gradient algorithm. Our model update frequently the policy of network switches in a way that, packet be forwarded to base stations with an optimized power level. The policies taken by the network controller are evaluated with a network simulator to ensure the energy consumption reduction and quality of service balance. The reinforcement learning model allows us to constantly learn and adapt to the changing situations in the dynamic network environment, hence having a more robust and realistic intelligent network management policy set. Our results demonstrate that energy efficiency can be enhanced by 32% and 67% in dense and sparse scenarios, respectively.Thesis (M.S.) -- Graduate School of Natural and Applied Sciences. Computer Engineering
- …
