Bilgisayar Mühendisliği Bölümü Koleksiyonu

Bu koleksiyon için kalıcı URI

https://hdl.handle.net/20.500.14034/34

Listeleniyor 1 - 20 / 43

Categorization of customer complaints in food industry using machine learning approaches
(2022) Kılınç, Deniz; Doğan, Onur; Bozyiğit, Fatma
Customer feedback is one of the most critical parameters that determine the market dynamics of product development. In this direction, analyzing product-related complaints helps sellers to identify the quality characteristics and consumer focus. There have been many studies conducted on the design of Machine Learning (ML) systems to address the causes of customer dissatisfaction. However, most of the research has been particularly performed on English. This paper contributes to developing an accurate categorization of customer complaints about package food products, written in Turkish. Accordingly, various ML algorithms using TF-IDF and word2vec feature representation strategies were performed to determine the category of complaints. Corresponding results of Linear Regression (LR), Naive Bayes (NB), k Nearest Neighbour (kNN), Support Vector Machine (SVM), Random Forest (RF), and Extreme Gradient Boosting (XGBoost) classifiers were provided in related sections. Experimental results show that the best-performing method is XGBoost with TF-IDF weighting scheme and it achieves %86 F-measure score. The other considerable point is word2vec based ML classifiers show poor performance in terms of F-measure compared to the TF-IDF term weighting scheme. It is also observed that each experimented TF-IDF based ML algorithm gives a more successful prediction performance on the optimal subsets of features selected by the Chi Square (CH2) method. Performing CH2 on TF-IDF features increases the F-measure score from 86% to 88% in XGBoost.
A comparative study on COVID-19 prediction using deep learning and machine learning algorithms: a case study on performance analysis
(2022) Arslan, Hilal; Er, Orhan
COVID-19 disease has been the most important disease recently and has affected serious number of people in the world. There is not proven treatment method yet and early diagnosis of COVID-19 is crucial to prevent spread of the disease. Laboratory data can be easily accessed in about 15 minutes, and cheaper than the cost of other COVID-19 detection methods such as CT imaging and RT-PCR test. In this study, we perform a comparative study for COVID-19 prediction using machine learning and deep learning algorithms from laboratory findings. For this purpose, nine different machine learning algorithms including different structures as well as deep neural network classifier are evaluated and compared. Experimental results conduct that cosine k-nearest neighbor classifier achieves better accuracy with 89% among other machine learning algorithms. Furthermore, deep neural network classifier achieves an accuracy of 90.3% when one hidden layer including 60 neurons is used to detect COVID-19 disease from laboratory findings data.
AFWDroid- Deep feature extraction and weighting for android malware detection
(2021) Er, Orhan; Arslan, Recep Sinan; Ölmez, Emre
Android malware detection is a critical and important problem that must be solved for a widely used operating system. Conventional machine learning techniques first extract some features from applications, then create classifiers to distinguish between malicious and benign applications. Most of the studies available today ignore the weighting of the obtained features. To overcome this problem, this study proposes a new software detection method based on weighting the data in feature vectors to be used in classification. To this end, firstly, the manifest file was read from the Android application package. Different features such as activities, services, permissions were extracted from the file, and for classification, a selection was made among these features. The parameters obtained as a result of selection were optimized by the deep neural network model. Studies revealed that through feature selection and weighting, better performance values could be achieved and more competitive results could be obtained in weight-sensitive classification.
Sequential rule mining for analysis of customer movements in different visits
(IEOM Society, 2021) Taşer, Pelin Yıldırım; Doğan, Onur
Developing technologies in customer analytics provide several opportunities to retailers. Analyzing customer movements in the stores as a part of customer analytics can reveal various shopping behaviors. It facilitates to understand better customers' visit purposes. This study applies four sequential mining algorithms, CMRules, CMDeo, ERMiner, and RuleGrowth, to analyze the visit purposes of customers in a supermarket. Moreover, it compares variations among different visits belonging to the same customers. This study concludes three main results. First, it indicates that customers prefer visiting the supermarket not only for their specific needs but also for all their needs at every visit. Second, the ERMiner algorithm is faster than the other algorithms. Third, customers who visit {Construction, Kitchen} and {Sanitary ware, Garden} bought at least one product with a high probability. Moreover, this study describes the concept of interesting rule, which has a lower support value and higher confidence value. Customers can visit the supermarket for various purposes resulting in different interesting rules. As an interesting rule in the second visit, purchased customers visited Construction, Garden and Kitchen aisles before leaving the supermarket whereas this rule did not appear in the first visits. Customers visited the Construction aisle more after they visited the Entrance and Ironmongery aisles in their second visit. © IEOM Society International.
Enterprise blockchain-based privacy sharing on internet of things devices
(Institute of Electrical and Electronics Engineers Inc., 2021) Öksüzer Şafak; Dalkılıç Gökhan; Kösemen Cem
Nowadays, the Internet of things systems became integrated into our lives. Data produced by the sensors on these systems can be considered personal data that is private. Legal regulations such as the general data protection regulation GDPR secure the storage and sharing of these personal data. It is not easy to automatically control these systems with legal regulations. In this study, we present a new privacy-sharing method of providing privacy for personal data sharing. This method ensures that using blockchain technology secures all privacy sharing steps. We used Quorum as the blockchain infrastructure that is an enterprise blockchain, making data transparently available to the peers in a private and permissioned network. © 2021 IEEE
Machine learning applications for Fraud Detection in finance sector
(Springer, 2022) Taşer, Pelin Yıldırım; Bozyiğit, Fatma
Due to advances in information technology, instantaneous accessibility to financial services through digital channels has increased. Although digital platforms’ usage makes an individual’s life more comfortable, it may also cause some critical consequences like financial fraud which causes critical losses for companies in the industrial sector, investors, and governments. Identification of frauds can be challenging task for a human because it may be necessary to analyse high volume data during long time periods. An alternative is to use financial data as a fraud detection tool to automatically classify fraudulent activities. Currently, there are many practical solutions for automatically detect frauds in the finance domain. In this chapter, we examined on three different fraud types (bank fraud, insurance fraud, and corporate fraud) in finance sector and reviewed the studies using machine learning methods to detect financial fraud in a detailed manner. The findings from this review show that most commonly applied algorithms for financial fraud detection are Decision Tree, Support Vector Machine (SVM), Artificial Neural Networks (ANN), and Random Forest and most of machine learning-based studies were performed in bank fraud field. This chapter also reveals that deep learning and ensemble-based machine learning applications has been frequently preferred in recent years to improve detection performance of the frauds in finance sector. © 2022, The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd.
Lib2Desc: automatic generation of security-centric Android app descriptions using third-party libraries
(Springer, 2022) Çevik, Beyza; Altıparmak, Nur; Aksu, Murat; Şen, Sevil
Android app developers are expected to specify the use of dangerous permissions in their app descriptions. The absence of such data indicates suspicious behavior. However, this is not always caused by the malicious intent of developers; it may be due to the lack of documentation of the third-party libraries they use. To fill this gap in the literature, this study aims to enrich application descriptions with security-centric information of third-party libraries. To automatically generate application definitions, the study explores classifying libraries and extracting code summaries of library methods that use dangerous permissions and/or leak data. Both the textual information of third-party libraries and their source code are used to create these definitions. To the best of our knowledge, this is the first approach in the literature that creates app descriptions based on third-party libraries.
ERIM: An ensemble of rare itemset mining and its application in the automotive industry
(Wiley, 2022) Akdaş, Devrim Naz; Birant, Derya; Taşer, Pelin Yıldırım
Discovering previously unknown anomalies that are rare and dramatically differ from the majority of the data is a critical need for the automotive industry. Rare itemset mining (RIM), one of the pattern-based methods, has been used for anomaly detection due to providing successful analysis results. However, several aspects still need to be explored, such as improving the mining process by identifying more targeted, valuable and reliable rare itemsets. Motivated by this fact, this study proposes a novel approach, named ensemble of rare itemset mining (ERIM), which investigates weak rare itemsets (WRIs) using different algorithms and aggregates these rules to obtain strong rare itemsets (SRIs). This study also combines four different RIM algorithms (Apriori Rare, Apriori Inverse, CORI and RP-Growth) as base learners for the first time. The proposed ERIM approach is a general methodology that can be applied to any field, but, in this study, it was used in the automotive industry as a case study. In the experiments, ERIM was applied to a real-world gear manufacturing dataset to discover anomalies in machine downtimes. The experimental results were evaluated in terms of the number of itemsets and the length of itemsets by giving some samples, as well. The results showed that the proposed ERIM approach gives more reliable common knowledge by jointly considering the relation between WRIs discovered by the base learners. The findings indicated that the proposed ERIM technique was successful in detecting anomalies whose support values are below 7.12. Furthermore, it is clear from the experimental results that the ERIM discovered the highest number of SRIs, 1403, each of which is a 3-itemset. Finally, the results showed that our method performed 43.37% better on average than state-of-the-art methods on the same dataset.
Comparison of experimental measurements and machine learning predictions of dielectric constant of liquid crystals
(Indian Acad Sciences, 2022) Taşer, Pelin Yıldırım; Önsal, Gülnur; Uğurlu, Onur
In this study, we investigated the dielectric properties of the phthalocyanine (Pc)-doped nematic liquid crystal (NLC) composite structures. 4-Pentyl-4 & PRIME;-cyanobiphenyl (5CB) NLC was dispersed with 1 and 3% wt/wt Pc to investigate the doping concentration effect. Dielectric measurements of the samples were carried out using the dielectric spectroscopy method. Moreover, the real and imaginary components of the dielectric constant values were estimated based on the input parameters (frequency, voltage value and dispersion rate) using two different traditional regression algorithms (k-Nearest Neighbor and Decision Tree Regression) and five different ensemble-based regression algorithms (Extreme Gradient Boosting, Random Forest, Extra Tree Regression, Voting and Bagging using k-Nearest Neighbor as a base learner). According to the obtained results, the Extra Tree Regression algorithm had the best prediction performance on real and imaginary components of the dielectric constant values. Moreover, it is seen from the obtained results that the ensemble-based regression algorithms are more successful than the traditional ones.
An ordinal multi-dimensional classification (OMDC) for predictive maintenance
(Tech Science Press, 2023) Taşer, Pelin Yıldırım
Predictive Maintenance is a type of condition-based maintenance that assesses the equipment's states and estimates its failure probability and when maintenance should be performed. Although machine learning techniques have been frequently implemented in this area, the existing studies disregard to the natural order between the target attribute values of the historical sensor data. Thus, these methods cause losing the inherent order of the data that positively affects the prediction performances. To deal with this problem, a novel approach, named Ordinal Multi-dimensional Classification (OMDC), is proposed for estimating the conditions of a hydraulic system's four components by taking into the natural order of class values. To demonstrate the prediction ability of the proposed approach, eleven different multi-dimensional classification algorithms (traditional Binary Relevance (BR), Classifier Chain (CC), Bayesian Classifier Chain (BCC), Monte Carlo Classifier Chain (MCC), Probabilistic Classifier Chain (PCC), Classifier Dependency Network (CDN), Classifier Trellis (CT), Classifier Dependency Trellis (CDT), Label Powerset (LP), Pruned Sets (PS), and Random k-Labelsets (RAKEL)) were implemented using the Ordinal Class Classifier (OCC) algorithm. Besides, seven different classification algorithms (Multilayer Perceptron (MLP), Support Vector Machine (SVM), k-Nearest Neighbour (kNN), Decision Tree (C4.5), Bagging, Random Forest (RF), and Adaptive Boosting (AdaBoost)) were chosen as base learners for the OCC algorithm. The experimental results present that the proposed OMDC approach using binary relevance multi-dimensional classification methods predicts the conditions of a hydraulic system's multiple components with high accuracy. Also, it is clearly seen from the results that the OMDC models that utilize ensemble-based classification algorithms give more reliable prediction performances with an average Hamming score of 0.853 than the others that use traditional algorithms as base learners.
Automatic keyword assignment system for medical research articles using nearest-neighbor searches
(Scientific And Technological Research Council Turkey, 2022) Dilmaç, Fatih; Alpkoçak, Adil
Assigning accurate keywords to research articles is increasingly important concern. Keywords should be selected meticulously to describe the article well since keywords play an important role in matching readers with research articles in order to reach a bigger audience. So, improper selection of keywords may result in less attraction to readers which results in degradation in its audience. Hence, we designed and developed an automatic keyword assignment system (AKAS) for research articles based on k-nearest neighbor (k-NN) and threshold-nearest neighbor (t-NN) accompanied with information retrieval systems (IRS), which is a corpus-based method by utilizing IRS using the Medline dataset in PubMed. First, AKAS accepts an abstract of the research article or a particular text as a query to the IRS. Next, the IRS returns a ranked list of articles to the given query. Then, we selected a set of documents from this list using two different methods, which are k-NN and t-NN representing the first k documents and documents whose similarity is greater than the threshold value of t, respectively. To evaluate our proposed system, we conducted a set of experiments on a selected subset of 458,594 PubMed articles. Then, we performed an experiment to observe the performance of AKAS results by comparing with the original keywords assigned by authors. The results we obtained showed that our system suggests keywords more than 55% match in terms of F-score. We presented both methods we used and results of experiments, in detail.
A novel machine learning approach: Soil temperature ordinal classification (STOC)
(Ankara Univ, Fac Agr, 2022) Kucuk, Cansel; Birant, Derya; Yıldırım Taşer, Pelin
Soil temperature prediction is an important task since soil temperature plays an important role in agriculture and land use. Although some progress has been made in this area, the existing methods provide a regression or nominal classification task. However, ordinal classification is yet to be explored. To bridge the gap, this paper proposes a novel approach: Soil Temperature Ordinal Classification (STOC), which considers the relationships between the class labels during soil temperature level prediction. To demonstrate the effectiveness of the proposed approach, the STOC method using five different traditional machine learning methods (Decision Tree, Naive Bayes, K-Nearest Neighbors, Support Vector Machines, and Random Forest) was applied on daily values of meteorological and soil data obtained from 16 stations in three states (Utah, Alabama, and New Mexico) of United States at five soil depths (2, 4, 8, 20, and 40 inches) between the years of 2011 and 2020. The experiments show that the proposed STOC approach is an efficient method for soil temperature level (very low, low, medium, high, and very high) prediction. The applied STOC models (STOC.DT, STOC.NB, STOC.KNN, STOC.SVM, and STOC.RF) showed average accuracy rates of 90.95%, 77.09%, 90.84%, 89.94%, and 90.91% on the experimental datasets, respectively. It was observed from the experimental results that the STOC.DT method achieved the best soil temperature level prediction among the others.
A novel multi-view ordinal classification approach for software bug prediction
(Wiley, 2022) Taşer, Pelin Yıldırım
Software bug prediction aims to enhance software quality and testing efficiency by constructing predictive classification models using code properties. This enables the prompt detection of fault-prone modules. There are several machine learning-based software bug prediction studies, which mainly focus on single view data by disregarding the natural ordering relation among the class labels in the literature. Thus, these studies cause losing each view's own intrinsic structure and the inherent order of the labels that positively affect the prediction performance. To overcome this drawback, this study focuses on integrating ordering information and a multi-view learning strategy. This paper proposes a novel approach multi-view ordinal classification (MVOC), which learns from different views (complexity, coupling, cohesion, inheritance and scale) of the software dataset separately and predicts software bugs taking the inherent order of class labels (non-buggy, less buggy and more buggy) into consideration. To demonstrate its prediction performance, the MVOC approach was executed on the 40 different real-world software datasets using six different classification algorithms as base learners. In the experiments, the MVOC approach was compared with traditional classifiers and their multi-view implementations in terms of precision, recall, f-measure and accuracy rate metrics. The results indicate that the MVOC approach presents better prediction performance on average than the multi-view-based and traditional classifiers. It is also observed from the results that the MVOC.RF model achieved the highest classification performance with an average accuracy rate of 85.65%.
A comparative study on prediction of survival event of heart failure patients using machine learning algorithms
(Springer London Ltd, 2022) Karakuş, Mücella Özbay; Er, Orhan
Cardiovascular diseases cause approximately 17 million deaths each year and 31% of deaths worldwide. These diseases generally occur as myocardial infarction and heart failure. The survival status, which we used as a target in our classification study, indicates that the patient died or survived before the end of the follow-up period, which is a mean of 130 days. Various machine learning classifiers have been preferred to both predict survival of patients and rank the characteristics corresponding to the most important risk factors. For this purpose, the data set that is occurred totally 299 samples is traditionally divided into 70% for training and 30% for test cluster to be used in machine learning algorithms, with have been analyzed with many methods such as Artificial Neural Networks, Fine Gaussian SVM, Fine KNN, Weighted KNN, Subspace KNN, Boosted Trees, and Bagged Trees. As a result, according to the data obtained, it has been seen that there are algorithms that can predict heart failure diagnosis with full accuracy (100%). Thus, it was concluded that it is appropriate to use machine learning algorithms to predict whether a heart failure patient will survive. This study has the potential to be used as a new supportive tool for doctors when predicting whether a heart failure patient will survive.
A quantitative evaluation of explainable AI methods using the depth of decision tree
(Scientific And Technological Research Council Turkey, 2022) Ahmed, Nizar Abdulaziz Mahyoub; Alpkoçak, Adil
It is necessary to develop an explainable model to clarify how and why a medical model makes a particular decision. Local posthoc explainable AI (XAI) techniques, such as SHAP and LIME, interpret classification system predictions by displaying the most important features and rules underlying any prediction locally. Therefore, in order to compare two or more XAI methods, they must first be evaluated qualitatively or quantitatively. This paper proposes quantitative XAI evaluation metrics that are not based on biased and subjective human judgment. On the other hand, it is dependent on the depth of the decision tree (DT) to automatically and effectively measure the complexity of XAI methods. Our study introduces a novel XAI strategy that measures the complexity of any XAI method by using a characteristic of another model as a proxy. The output of XAI methods, specifically feature importance scores from SHAP and LIME, is fed into the DT in our proposal. The DT will then draw a full tree based on the feature importance score decisions. As a result, we developed two main metrics that can be used to assess the DT's complexity and thus the associated XAI method: the total depth of the tree (TDT) and the average of the weighted class depth (ACD). The results show that SHAP outperforms LIME and is thus less complex. Furthermore, in terms of the number of documents and features, SHAP is more scalable. These results can indicate whether a specific XAI method is suitable for dealing with different document scales. Furthermore, they can demonstrate which features can be used to improve the performance of the black-box model, in this case, a feedforward neural network (FNN).
Effectiveness of social media in stock market price prediction based on machine learning
(Springer International Publishing Ag, 2022) Karaşahin, Emre; Utku, Semih; Öztürkmenoğlu, Okan
Trying to predict the future using social media data and analytics is very popular today. With this motivation, we aimed to make stock market predictions by creating different analysis models for 10 different banks traded in Borsa Istanbul 100 over 3 different groups that we selected on social media. The groups determined within the scope of the study can be detailed as tweets posted by banks from their accounts, tweets posted with the name of the bank, and tweets with the name of the bank posted from approved accounts. In our analysis, we used various variations, including the tweets' sentiments, replies, retweet and like counts of the tweets, the effects of daily currency (Dollar, Euro, and Gold) prices, and the changes in stock changes up to 3 days. To apply some pre-processing techniques to the collected data, we defined sentiment classes for sentiment analysis, created 6 different models, and analyzed it using 7 different classification algorithms such as Multi-Layer Perceptron, Random Forest, and deep learning algorithm. After all the models and analysis, we got a total of 1440 different results. According to our results, the accuracy rates vary according to the data groups and models we have chosen. The tweet group in which the name of the banks is mentioned can be shown as the most successful data group and we can easily say that there is a certain relation between social media and stock market prices.
Multiple-classifiers in software quality engineering: Combining predictors to improve software fault prediction ability
(Elsevier - Division Reed Elsevier India Pvt Ltd, 2020) Yücalar, Fatih; Özçift, Akın; Borandağ, Emin; Kılınç, Deniz
Software development projects require a critical and costly testing phase to investigate efficiency of the resultant product. As the size and complexity of project increases, manual prediction of software defects becomes a time consuming and costly task. An alternative to manual defect prediction is the use of automated predictors to focus on faulty modules and let the software engineer to examine the defective part with more detail. In this aspect, improved fault predictors will always find a software quality application project to be applied on. There are many base predictors tested-designed for this purpose. However, base predictors might be combined with an ensemble strategy to further improve to increase their performance, particularly fault-detection abilities. The aim of this study is to demonstrate fault-prediction performance of ten ensemble predictors compared to baseline predictors empirically. In our experiments, we used 15 software projects from PROMISE repository and we evaluated the fault-detection performance of algorithms in terms of F-measure (FM) and Area under the Receiver Operating Characteristics (ROC) Curve (AUC). The results of experiments demonstrated that ensemble predictors might improve fault detection performance to some extent. (C) 2019 Karabuk University. Publishing services by Elsevier B.V.
A real-world text classification application for an e-commerce platform
(IEEE, 2019) Yıldırım, Fatih Mehmet; Kaya, Abdullah; Öztürk, Selin Nur; Kılınç, Deniz
This study aims to reflect the real-world utilities of machine learning applications by implementing a set of different text classification algorithms in terms of accuracy and performance. We developed a system capable of executing various text classification algorithms and generated models trained on real product catalog data collected from morhipo.com, an online fashion commerce platform. The highest mean accuracy rate was obtained as 96.08% (ranging between 85.44% and 99.99%) with a standard deviation value of 5.65% by Linear Support Vector Classifier (LinearSVC) algorithm.
Betweenness centrality in sparse real world and wireless multi-hop networks
(Springer Science and Business Media Deutschland GmbH, 2022) Tuzcu, Atakan; Arslan, Hilal
Graphs are one of the compact ways to represent information about real-life and intelligent system networks like wireless sensor networks. Betweenness centrality is an important network measure that evaluates the significance of a node based on the shortest paths and is widely used in biological, social, transportation, complex, and communication networks. In this study, we implement an efficient algorithm computing betweenness centrality of nodes for real-world and wireless multi-hop networks. Large sparse graphs are stored using compressed sparse row storage format and modified version of Dijkstra’s algorithm is used to compute shortest paths. We conduct a comprehensive experimental study on real-world networks as well as wireless sensor networks that are state-of-the-art technologies for different applications such as intelligence structures, industrial and home automation as well as health care. We evaluate the effect of network dimension on the time needed to compute betweenness centrality. Experimental results demonstrate that computation time required to compute betweenness centrality varies from 0.9 to 52.5 s when the number of vertices changes from 10,000 to 60,000. We also observe that the proposed algorithm efficiently computes betweenness centrality for networks coming from machine learning, power network, and networks obtained from optimization problems as well as computational fluid dynamics. © 2022, The Author(s), under exclusive license to Springer Nature Switzerland AG.
Kapalı ortamlarda kişi tespitinde makine öğrenmesi algoritmalarının karşılaştırmalı başarım analizi
(2021) Taşer, Pelin Yıldırım; Akram, Vahid Khalilpour
Günümüzde, iç mekan konumlandırma ve kişi takip sistemlerinin uygulama alanları her geçen gün artış göstermektedir. Özellikle, hasta, personel, cihaz ve müşteri takip sistemleri ile akıllı binalar ve kalabalık tahminleme gibi birçok alanda, kişilerin konumlarının veya mekan içerisinde bulunma durumlarının doğru tespiti büyük önem taşımaktadır. İç mekan konumlandırma sistemlerinde genellikle, hedef mobil varlığın üzerine periyodik olarak radyo sinyali gönderen küçük bir cihaz yerleştirilir ve bu cihazdan elde edilen sinyaller ile varlığın konumu belirlenir. Fakat bazı ortamlarda, üzerinde herhangi bir sinyal göndericisi taşımayan varlıkların konumlarının tespit edilmesine ihtiyaç duyulmaktadır. Dolayısıyla, mobil cihaz kullanılmayan radyo tabanlı takip sistemlerinde, radyo sinyallerinde meydana gelen dalgalanmalar analiz edilerek, ortamdaki hareketlilik tahmin edilmeye çalışılır. Bu sistemlerde, ortamın çeşitli noktalarına periyodik olarak radyo sinyalleri gönderen ve diğer cihazların gönderdiği sinyalleri alabilen cihazlar yerleştirilir. Ortamda bulunan herhangi bir nesnenin hareket etmesi durumunda, sinyal gücündeki dalgalanmalar analiz edilerek, ortamdaki hareketlilik ve yoğunluk tahmin edilebilir. Ancak bazı durumlarda, radyo sinyallerinde hareketten kaynaklanmayan, geçici ama nispeten şiddetli dalgalanmalar yaşanabilmektedir. Bu tür dalgalanmalar, yanlış tespitlere sebep olarak, sistemin hassasiyetini ve doğruluğunu düşürmektedir. Makine öğrenmesi tekniklerinin, veriler arasındaki gizli örüntü ve karmaşık ilişkileri ortaya çıkarmadaki başarıları sayesinde, makine öğrenmesi tekniklerine dayalı kişi tespit sistemleri, geleneksel yöntemlere göre doğruluğu daha yüksek tahminleme becerisi sunmaktadır. Dolayısıyla, bu çalışmada, kapalı ortamda kişi tespiti için makine öğrenmesi algoritmalarından yararlanılmıştır. Deneysel çalışmalar kapsamında, 10 farklı geleneksel (Naive Bayes, Çok Katmanlı Algılayıcı (MLP), Destek Vektör Makineleri (SVM) ve K-en Yakın Komşuluk (K-NN)) ve karar ağacı tabanlı (C4.5, Random Forest, Random Tree, REPTree, Decision Stump ve HoeffdingTree) sınıflandırma algoritmaları, kapalı ortamdaki 3 farklı telsiz duyarga düğümünden elde edilen ve 23585 kayıttan oluşan veri seti üzerinde ayrı ayrı uygulanmış, doğruluk oranı ve model oluşturma süresi performanslarına göre karşılaştırılmıştır. Elde edilen sonuçlar incelendiğinde, çalışmada uygulanan tüm algoritmaların kapalı alandaki kişi tespitinde %80’in üzerinde başarı performansı sunduğu ve en başarılı algoritmanın %99.68 doğruluk oranı ile Random Forest olduğu gözlemlenmiştir. Ayrıca, geleneksel ve karar ağacı tabanlı sınıflandırma algoritmaları sağlamış oldukları ortalama doğruluk oranlarına göre kıyaslandığında ise karar ağacı tabanlı algoritmaların %95.78 ile daha yüksek tahminleme becerisi sunduğu görülmektedir.

Güncel Gönderiler