Tülin İnkaya’s research while affiliated with Bursa Uludağ Üni̇versi̇tesi̇ and other places

Dijital dönüşüm ile tedarik zinciri yönetiminde büyük veri analitiğinin önemi gün geçtikçe artmaktadır. Özellikle müşteri taleplerinin hızlı ve doğru tahmin edilmesinde büyük verinin kullanımı firmalara rekabet avantajı sağlamaktadır. Bu doğrultuda, yapay zekâ tekniklerinden biri olan derin öğrenme modelleri büyük verideki karmaşık örüntülerin keşfedilmesinde öne çıkmaktadır. Son yıllarda literatürde çok sayıda derin öğrenme yöntemi önerilmiştir. Bu çalışmada, satış tahmini problemi için derin öğrenme yöntemlerinin performansları karşılaştırılmıştır. Bu kapsamda derin sinir ağı (DNN), derin otokodlayıcı (Deep AE), evrişimli sinir ağı (CNN), tekrarlayan sinir ağı (RNN), uzun kısa-süreli bellek (LSTM) ağı, çift yönlü LSTM (Bi-LSTM) ağı, kapılı tekrarlayan birim (GRU), CNN-LSTM ve evrişimli LSTM (ConvLSTM) yöntemleri uygulanmıştır. Çeşitli sektörlere ait satış verileri kullanılarak deneysel çalışmalar gerçekleştirilmiştir. Hiperparametre optimizasyonu ardından ele alınan yöntemlerin performansları tahmin doğruluğu ve eğitim süreleri açısından karşılaştırılarak sonuçların istatistiksel anlamlılığı değerlendirilmiştir. Sonuç olarak, LSTM ve GRU modellerinin tahmin doğruluğunda başarılı sonuçlar verdiği, CNN modelinin ise eğitim süresini kısalttığı görülmüştür.

K-nearest neighbor (KNN) algorithm is a widely used machine learning technique for prediction problems due its simplicity, flexibility and interpretability. When predicting the output variable of a data point, it basically averages the output values of its k closest neighbors. However, the impact of the neighboring points on the estimation may differ. Even though there are weighted versions of KNN, the effect of outliers and density differences within the neighborhoods are not considered. In order to fill this gap, we propose a novel weighting scheme for KNN regression based on local outlier factor (LOF). In particular, we combine the inverse of the Euclidean distance and LOF value so that the weights of the neighbors are determined using not only distance and connectivity but also outlier and density information around the neighborhood. Also, bootstrap aggregation is used to leverage the stability and accuracy of the LOF weighted KNN regression. Using real-life benchmark datasets, extensive experiments and statistical tests were performed for evaluating the performance of the proposed approach. The experimental results indicate the superior performance of the proposed approach in small neighborhood sizes. Moreover, the proposed approach was implemented in a make-to-order manufacturing company, and die production times were estimated successfully.

Accurate remaining useful life (RUL) prediction is one of the most challenging problems in the prognostics of turbofan engines. Recently, RUL prediction methods for turbofan engines mainly involve data-driven models. Preprocessing the sensor data is essential for the performance of the prognostic models. Most studies on turbofan engines use piecewise linear (PwL) labeling, which starts with a constant initial RUL value in normal/healthy operating time. In this study, we designed a prognostic procedure that includes difference-based feature construction, change-point-detection-based PwL labeling, and a 1D-CNN-LSTM (one-dimensional–convolutional neural network–long short-term memory) hybrid neural network model for RUL prediction. The procedure was evaluated on the subset FD001 of the C-MAPSS dataset. The proposed procedure was compared with machine learning and deep learning models with and without the new difference feature. Also, the results were compared with the studies that used similar labeling approaches. Our analysis of the numerical results underscores the clear superiority of the proposed 1D-CNN-LSTM model with the difference feature in RUL prediction, with a score of 437.2 and an RMSE value of 16.1. This result illustrates the superior predictive capability of the 1D-CNN-LSTM model, which outperformed traditional machine learning methods and one of the earliest deep learning methods. These findings emphasize the superior predictive capability of the 1D-CNN-LSTM model and underline the potential of the feature engineering process for more accurate and robust RUL prediction in the context of turbofan engine prognostics.

The increase of digitalization has enhanced the collection of time series data using sensors in various production and service systems such as manufacturing, energy, transportation, and healthcare systems. To manage these systems efficiently and effectively, artificial intelligence techniques are widely used in making predictions and inferences from time series data. Artificial intelligence methods require a sufficient amount of labeled data in the learning process. However, most of the data in real-life systems are unlabeled, and the annotation task is costly or difficult. For this purpose, active learning can be used as a solution approach. Active learning is one of the machine learning methods, in which the model interacts with the environment and requests the labels of the informative samples. In this study, we introduce an active learning-based approach for the time series classification problem. In the proposed approach, the k-medoids clustering method is first used to determine the representative samples in the dataset, and these cluster representatives are labeled during the initialization of active learning. Then, the k-nearest-neighbor (KNN) algorithm is used for the classification task. For the query selection, uncertainty sampling is applied so that the samples having the least certain labels are prioritized. The performance of the proposed approach was evaluated using sensor data from the production and healthcare systems. In the experimental study, the impacts of the initialization techniques, number of queries, and neighborhood size were analyzed. The experimental studies showed the promising performance of the proposed approach compared to the competing approaches.

With the emergence of digitilization, numeric prediction has become a prominent problem in various fields including finance, engineering, industry, and medicine. Among several machine learning methods, regression tree is a widely preferred method due to its simplicity, interpretability and robustness. Motivated by this, we introduce a novel ensemble regression tree based methodology, namely LOF-BRT+OR. The proposed methodology is an integrated solution approach with outlier removal, regression tree and ensemble learning. First, irregular data points are removed using local outlier factor (LOF), which measures the degree of being an outlier for each point. Next, a novel regression tree with LOF weighted node model is introduced. In the proposed node model, the weights of the points in the nodes are determined according to their surrounding neighborhood, as a function of LOF values and neighbor ranks. Finally, in order to increase the prediction performance, ensemble learning is adopted. In particular, bootstrap aggregation is used to generate multiple regression trees with LOF weighted node model. The experimental study shows that the proposed methodology yields the best root mean squared error (RMSE) values in five out of nine data sets. Also, the non-parametric tests demonstrate the statistical significance of the proposed approach over the benchmark methods. The proposed methodology can be applicable to various prediction problems.

Üretim ve hizmet sektörlerinde faaliyet gösteren firmalar, artan rekabet koşulları ile mücadele edebilmek için belirsizlik altında geleceğe yönelik çeşitli kararlar alırlar. Bu kritik kararlardan biri satış tahminidir. Dijital teknolojilerin yaygınlaşması ile derin öğrenme yaklaşımlarının satış tahmininde kullanımı artmaktadır. Derin öğrenme, başarılı sonuçlar vermesine rağmen büyük miktarda veri ile uzun eğitim sürelerine ihtiyaç duymaktadır. Bu duruma çözüm olarak problemler arası bilgi aktarımını sağlayan transfer öğrenme (TL) kullanılmaktadır. Transfer öğrenme, kaynak veriler ile modelin eğitimini ve hedef veriye aktarımını sağlamaktadır. Bu çalışmada, farklı ürünlerin satış tahmini modellerinden elde edilen bilginin gelecekteki tahmin modellerine aktarımını sağlamak üzere derin transfer öğrenme yaklaşımı önerilmiştir. Satış verisi tek değişkenli zaman serisi olarak ele alınmıştır. Kaynak veri seçiminde aktarılabilirlik ölçütü olarak hedef ve kaynak veri arasındaki gerçek cezalı düzenleme uzaklığı (ERP) kullanılmıştır. Seçilen kaynak veri ile zamansal bağımlılıkların modellenmesini sağlayan uzun kısa vadeli hafıza (LSTM) ağı eğitilmiştir. Ön eğitilen LSTM ağında parametre transferi yapılarak hedef veri için ERP-LSTM-TL tahmin modeli oluşturulmuştur. Çeşitli sektörlere ait satış veri kümelerinde yapılan deneysel çalışmalarda ERP-LSTM-TL, hedef veri ile eğitilen LSTM’e göre tahmin doğruluğunda ve eğitim süresinde iyileşme sağlamıştır. Önerilen yaklaşımın performansı klasik tahmin ve makine öğrenmesi yöntemlerinin performansları ile karşılaştırılmıştır. ERP-LSTM-TL karşılaştırılan yöntemlere göre istatistiksel olarak daha iyi sonuç vermiştir.

A similarity graph represents the local characteristics of a data set, and it is used as input to various clustering methods including spectral, graph-based, and hierarchical clustering. Several similarity graphs exist in the literature; however, there is not a single similarity graph that can handle all kinds of cluster shapes and structures. In this study, motivated by the successful applications of ensemble approaches to clustering, a generic method for consensus similarity graph construction is proposed. The proposed approach first constructs multiple similarity graphs using bootstrap aggregating (bagging). Then, these graphs are fused into a consensus similarity graph using the normalized co-association matrix. We use k-nearest neighbor, ε$\varepsilon$-neighborhood, fully connected graph, and proximity graphs as the base similarity graphs. Moreover, the proposed approach is coupled with various clustering algorithms including spectral, graph-based, and hierarchical clustering. The experimental results with various spatial and real data sets demonstrate the effectiveness of the consensus similarity graphs in clustering. The proposed approach is also robust to local noise.

In machine learning, nearest neighbor (NN) regression is one of the most prominent methods for numeric prediction. It estimates the output variable of a new data point by averaging the output variables of the neighboring points. The selection of the neighborhood and its parameter(s) is crucial for the performance of NN regression, however this is still an open issue. This study contributes to the literature by adopting the parameter-free surrounding neighborhood (PSN) concept for NN regression. PSNs are based on proximity graphs, i.e. minimum spanning tree, relative neighborhood graph, and Gabriel graph. They yield a unique neighborhood for each point by combining proximity, connectivity and spatial distribution. The performances of the PSN regression methods are compared with k-nearest neighbors, k-nearest centroid neighbors, and support vector regression using real-world data sets. The statistical tests show that the PSN regression methods perform significantly better than most of the competing approaches. Also, the proposed approaches do not have any parameters to be set.