Satış ve talep tahmini için derin transfer öğrenme metodolojisinin geliştirilmesi
Date
2024-03-27
Authors
Erol, Begüm
Journal Title
Journal ISSN
Volume Title
Publisher
Bursa Uludağ Üniversitesi
Abstract
Satış ve talep tahmini, tedarik zinciri yönetiminde kritik bir role sahiptir. Hızlı, doğru ve etkin tahminler için gelişmiş tekniklere ihtiyaç duyulmaktadır. Gelişen teknoloji ile birlikte son yıllarda satış ve talep tahmininde derin öğrenme (Deep Learning – DL) yöntemlerinin kullanımı yaygınlaşmıştır. DL ile yapılan tahminlerde başarılı sonuçlar elde edilmektedir. Ancak DL eğitimi için hem büyük miktarda veriye ihtiyaç duyulmaktadır, hem de kullanılan veri miktarı arttıkça modelin eğitim süresi katlanarak artmaktadır. Buna çözüm olarak, kaynak alandaki veri ile ön eğitilmiş ağdan edinilen bilgilerin hedef alana aktarımını sağlayan transfer öğrenme kullanılmaktadır. Bu tez çalışmasında, satış ve talep tahmini için bilgi aktarımını sağlayan derin transfer öğrenme (Deep Transfer Learning – DTL) tabanlı metodolojiler geliştirilmiştir. Satış verileri tek ve çok değişkenli zaman serileri olarak ele alınmıştır. İlk olarak, satış ve talep tahmininde en uygun DL yöntemini seçmek amacıyla dokuz farklı DL yöntemi karşılaştırılmıştır. Parametrik olmayan istatistiksel testler yardımıyla LSTM yöntemi seçilmiştir. İkinci olarak, uzaklık tabanlı kaynak seçimi ile tek kaynaklı DTL modeli önerilmiştir. Sonrasında, transfer öğrenmede aktarım başarısını tahmin eden rastgele orman modeli geliştirilerek kaynak seçiminde açıklanabilirlik sağlanmıştır. Rastgele ormanda seçilen öznitelikler ile derin kümeleme tabanlı DTL modeli geliştirilmiştir ve bu modele topluluk öğrenme uygulanmıştır. Bu metodoloji kümeleme ve topluluk tabanlı DTL metodolojisi olarak adlandırılmıştır. Ayrıca, geliştirilen metodolojide çoklu kaynak seçimi için ağırlıklandırma mekanizması geliştirilmiştir. Son olarak, önerilen metodolojiler çok değişkenli satış verilerine uyarlanmıştır. Çeşitli veri setleri ile yapılan deneysel çalışmalarda, kümelemenin benzer satış verilerinin tespitini sağlayarak tahmin doğruluğunu arttırdığı görülmüştür. Ayrıca, çok kaynaklı DTL yaklaşımlarının tek kaynaklı DTL yaklaşımlarına göre tahmin sonuçlarını iyileştirdiği gözlenmiştir. Geliştirilen metodolojiler; enerji, finans, çevre ve tedarik zinciri gibi çeşitli üretim ve hizmet sektörlerinde uygulanabilir niteliktedir.
Sales and demand forecasting plays a critical role in supply chain management. Advanced techniques are needed for fast, accurate and effective predictions. With the developing technology, the use of deep learning (DL) methods in sales and demand forecasting has become widespread in recent years. Successful results are obtained in predictions made with DL. However, a large amount of data is needed for DL training, and as the amount of data used increases, the training time of the model increases exponentially. As a solution to this, transfer learning is used, which transfers the data in the source area and the information obtained from the pre-trained network to the target area. In this thesis study, deep transfer learning (DTL) based methodologies have been developed that provide information transfer for sales and demand forecasting. Sales data are treated as univariate and multivariate time series. First, nine different DL methods were compared in order to choose the most appropriate DL method for sales and demand forecasting. The LSTM method was chosen with the help of non-parametric statistical tests. Secondly, a single-source DTL model with distance-based source selection is proposed. Subsequently, a random forest model that predicts transfer success in transfer learning was developed to ensure explainability in source selection. A deep clustering-based DTL model was developed with features selected in random forest and ensemble learning was applied to this model. This methodology has been called clustering and ensemble-based DTL methodology. Additionally, a weighting mechanism for multiple source selection has been developed in the developed methodology. Finally, the proposed methodologies are adapted to multivariate sales data. In experimental studies conducted with various data sets, it has been shown that clustering increases forecast accuracy by enabling the detection of similar sales data. Additionally, it has been observed that multi-source DTL approaches improve prediction results compared to single-source DTL approaches. Developed methodologies; it is applicable in various production and service sectors such as energy, finance, environment and supply chain.
Sales and demand forecasting plays a critical role in supply chain management. Advanced techniques are needed for fast, accurate and effective predictions. With the developing technology, the use of deep learning (DL) methods in sales and demand forecasting has become widespread in recent years. Successful results are obtained in predictions made with DL. However, a large amount of data is needed for DL training, and as the amount of data used increases, the training time of the model increases exponentially. As a solution to this, transfer learning is used, which transfers the data in the source area and the information obtained from the pre-trained network to the target area. In this thesis study, deep transfer learning (DTL) based methodologies have been developed that provide information transfer for sales and demand forecasting. Sales data are treated as univariate and multivariate time series. First, nine different DL methods were compared in order to choose the most appropriate DL method for sales and demand forecasting. The LSTM method was chosen with the help of non-parametric statistical tests. Secondly, a single-source DTL model with distance-based source selection is proposed. Subsequently, a random forest model that predicts transfer success in transfer learning was developed to ensure explainability in source selection. A deep clustering-based DTL model was developed with features selected in random forest and ensemble learning was applied to this model. This methodology has been called clustering and ensemble-based DTL methodology. Additionally, a weighting mechanism for multiple source selection has been developed in the developed methodology. Finally, the proposed methodologies are adapted to multivariate sales data. In experimental studies conducted with various data sets, it has been shown that clustering increases forecast accuracy by enabling the detection of similar sales data. Additionally, it has been observed that multi-source DTL approaches improve prediction results compared to single-source DTL approaches. Developed methodologies; it is applicable in various production and service sectors such as energy, finance, environment and supply chain.
Description
Keywords
Satış ve talep tahmini, Derin öğrenme, Transfer öğrenme, Kaynak seçimi, Derin kümeleme, Topluluk öğrenme, Açıklanabilirlik, Yapay zeka, Artificial intelligence, Sales and demand forecasting, Deep learning, Transfer learning, Source selection, Deep clustering, Ensemble learning, Explainability