Transformer tabanlı öznitelik bütünleştirme ile eksik veriye dayanıklı tahminsel model geliştirilmesi
Development of a predictive model robust to missing data via transformer-based feature integration
- Tez No: 1018514
- Danışmanlar: DR. ÖĞR. ÜYESİ MUHAMMED KOTAN
- Tez Türü: Yüksek Lisans
- Konular: Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve Kontrol, Computer Engineering and Computer Science and Control
- Anahtar Kelimeler: Belirtilmemiş.
- Yıl: 2026
- Dil: Türkçe
- Üniversite: Sakarya Üniversitesi
- Enstitü: Fen Bilimleri Enstitüsü
- Ana Bilim Dalı: Bilişim Sistemleri Mühendisliği Ana Bilim Dalı
- Bilim Dalı: Belirtilmemiş.
- Sayfa Sayısı: Belirtilmemiş.
Özet
Günümüz dijitalleşmiş iş ortamlarında satış tahmini, yalnızca geçmiş satış değerlerine dayalı zaman serisi modelleriyle ele alınamayacak kadar karmaşık bir hâl almıştır. Özellikle perakende ve hizmet sektörlerinde satış performansı; dijital etkileşimler, kullanıcı davranışları ve ürün bağlamı gibi çok sayıda heterojen veri kaynağından etkilenmektedir. Ancak bu dijital akışlar çoğu zaman sistematik ve yüksek oranlı eksiklikler içermekte, bu durum klasik tahmin modellerinin temsil gücünü önemli ölçüde sınırlandırmaktadır. Geleneksel yaklaşımlar eksik veriyi çoğunlukla ön işleme aşamasında doldurulması gereken bir sorun olarak ele alırken, eksikliğin davranışsal bir sinyal olabileceği gerçeğini göz ardı etmektedir. Bu tez çalışmasında, çok kaynaklı ve eksiklik içeren iş verilerinde satış tahmini performansını artırmak amacıyla, eksikliğe duyarlı Transformer tabanlı bir özellik entegrasyon çerçevesi önerilmiştir. Çalışma kapsamında eksik veriyi sıfır ile dolduran temel (Baseline) mimari, akışları ayrı ayrı işleyen Çift Akışlı (Dual-Stream) mimari, eksikliğe göre ölçekleme yapan Uyarlamalı (Adaptive) mimari ve önerilen Füzyon (Fusion) mimarisi olmak üzere dört farklı strateji geliştirilmiştir. Özellikle Füzyon mimarisinde, satış akışı ve dijital etkileşim akışı birbirinden yapısal olarak ayrıştırılarak, her biri için uzman Transformer kodlayıcılar tasarlanmış; bu kodlayıcıların çıktıları öğrenilebilir bir geçitleme (gating) mekanizması aracılığıyla dinamik olarak birleştirilmiştir. Böylece model, her bir gözlem için hangi veri akışının daha bilgilendirici olduğunu eksiklik desenlerine bağlı olarak otomatik biçimde öğrenebilmektedir. Önerilen mimariler, gerçek bir işletme ortamından elde edilen, 6.728 gözlemden oluşan ve dijital etkileşim verilerinde yaklaşık %47,4 oranında sistematik eksiklik barındıran bir veri seti üzerinde değerlendirilmiştir. 5-katmanlı çapraz doğrulama ile yapılan deneysel analizlerde, önerilen Füzyon mimarisi 0,90 ± 0,01 skoru, %27,3 Ortalama Mutlak Yüzde Hata (MAPE) ve 15,21 Ortalama Mutlak Hata (MAE) değerleri ile diğer tüm Transformer tabanlı yaklaşımları geride bırakmıştır. Ayrıca, güçlü bir tabular temel model olan XGBoost ile yapılan karşılaştırmalarda; XGBoost tam veri içeren örneklerde marjinal bir üstünlük sağlasa da, dijital etkileşim verisinin eksik olduğu kritik senaryolarda Füzyon mimarisinin istatistiksel olarak anlamlı düzeyde daha yüksek performans (Füzyon: 0,9182 vs. XGBoost: 0,8971) sergilediği tespit edilmiştir. Bu çalışma, eksik veriyi bir ön işleme problemi olarak değil, doğrudan öğrenme sürecinin parçası olarak ele alan bir yaklaşım sunmakta; çok kaynaklı iş verilerinde Transformer tabanlı mimarilerin etkinliğini hem teorik hem de deneysel düzeyde ortaya koymaktadır.
Özet (Çeviri)
In contemporary data-driven business ecosystems, the digitalization of consumer interactions has not merely expanded the volume of available data but has fundamentally reconfigured the epistemological foundations of predictive analytics. Sales forecasting, traditionally conceptualized as a univariate or moderately multivariate time-series problem grounded in historical transaction records, has evolved into a structurally heterogeneous and asymmetrically informed multi-source learning task. In digitally transformed retail and service environments, transactional Point-of-Sale (POS) data is increasingly complemented by behavioral traces generated through mobile applications, QR-based digital menus, web interfaces, and loyalty infrastructures. These auxiliary streams provide granular signals regarding user attention, browsing patterns, product visibility, and engagement intensity. While such augmentation enriches contextual representation and potentially enhances predictive performance, it simultaneously introduces a methodological tension rooted in structural asymmetry and data incompleteness across heterogeneous streams. In practical business environments, transactional sales data typically remains complete, stable, and operationally mandatory, whereas digital interaction streams are often optional, behaviorally mediated, and systematically incomplete. The resulting missingness patterns are neither random nor purely technical. Rather, they are frequently the consequence of user decisions, such as bypassing a digital interface or interacting through alternative channels. Within the statistical taxonomy of missing data mechanisms, such conditions correspond to the Missing Not At Random (MNAR) category, wherein the probability of missingness is intrinsically related to latent behavioral states or unobserved contextual variables. Under MNAR settings, the absence of data may encode meaningful information about user intent, interaction preference, channel substitution, or cognitive engagement. Consequently, missingness should not be treated as a stochastic artifact detached from the prediction target but as a potentially informative component of the data-generating process itself. Despite this conceptual reality, conventional machine learning pipelines predominantly address missing data through preprocessing heuristics such as mean substitution, k-nearest neighbor imputation, regression-based reconstruction, or case-wise deletion. These strategies implicitly assume that the missingness mechanism can be neutralized prior to model training and that the predictive algorithm need not account for data availability structure. While such assumptions may be tenable under Missing Completely At Random (MCAR) or certain Missing At Random (MAR) regimes, they are theoretically fragile in MNAR environments, where missingness is behaviorally endogenous. In digitally mediated business systems, the non-occurrence of a behavioral event—such as the absence of a product view—may carry predictive significance comparable to its occurrence. Therefore, imputation-based normalization risks attenuating signal reliability, distorting feature representations, and conflating behavioral absence with numerical neutrality. This thesis addresses this methodological gap by proposing a missingness-aware feature integration framework grounded in the Transformer architecture, explicitly designed for sales forecasting under systematic multi-source asymmetry. Rather than reconstructing missing values, the proposed approach operationalizes missingness as an adaptive control variable that regulates information flow across heterogeneous data streams. The forecasting task is formulated as a supervised regression problem defined over three interdependent components: a complete sales feature matrix \operatorname{X}_s, a sparse digital interaction matrix \operatorname{X}_v, and a binary mask vector M\in{0,1}^nencoding the availability structure of the digital stream at the observation level. This formulation enables the model to explicitly condition representation learning on data availability rather than treating missingness as a latent nuisance. To systematically examine integration strategies under MNAR conditions, four Transformer-based architectures were designed and comparatively evaluated. The Baseline configuration performs zero-imputation on missing digital features and processes concatenated sales and digital inputs through a unified Transformer encoder, effectively assuming that numerical placeholder substitution renders missingness ignorable. The Dual-Stream architecture maintains semantic separation between sales and digital streams through independent projection layers and parallel processing paths before combining representations via fixed weighted summation. While this design preserves structural distinction, it lacks adaptive stream-level modulation and assumes a static contribution ratio across observations. The Adaptive architecture introduces a learnable global scaling parameter that adjusts digital feature influence under missing-data scenarios; however, this adjustment operates at an aggregate level and does not provide instance-specific sensitivity to heterogeneous availability patterns. The principal methodological contribution of this study is the proposed Fusion Transformer architecture, conceptually inspired by the Mixture-of-Experts paradigm and reliability-aware modeling principles. The model incorporates two specialized Transformer encoders: a Sales-Only encoder trained exclusively on the complete transactional stream to provide a stable and low-variance baseline representation, and a Full-Data encoder trained on the joint sales and digital streams to capture enriched behavioral dependencies when auxiliary signals are present. These expert representations are integrated through a learnable gating mechanism implemented as a Multi-Layer Perceptron conditioned on categorical embeddings and the mask vector M. The gating network produces an observation-specific coefficient g\in[0,1], enabling a dynamic convex combination of expert outputs. In scenarios where digital interaction data is absent, the gating mechanism systematically increases reliance on the Sales-Only encoder, thereby interpreting missingness as an indicator of reduced auxiliary reliability. Conversely, when digital data is available, the model shifts emphasis toward the Full-Data encoder to leverage behavioral enrichment. Through this mechanism, missingness is transformed from a preprocessing obstacle into an explicit regulator of expert specialization and stream-level trust allocation. Empirical validation was conducted using a real-world dataset comprising 6,728 product-level observations collected from a commercial enterprise operating with a QR-based digital menu system. The dataset includes sales quantities, pricing variables, categorical descriptors, and digital interaction metrics, with approximately 47.4% systematic missingness in the digital stream, consistent with MNAR dynamics. To ensure robust generalization assessment, a stratified five-fold cross-validation protocol was employed, preserving product-category heterogeneity across folds and mitigating distributional bias. Hyperparameters were optimized via randomized search to ensure equitable architectural comparison and to avoid overfitting artifacts. Experimental findings demonstrate that the Fusion architecture consistently achieves superior predictive performance relative to the Baseline, Dual-Stream, and Adaptive variants. The model attains an \operatorname{R}^2score of 0.90 ± 0.01, a Mean Absolute Percentage Error (MAPE) of 27.3%, and a Mean Absolute Error (MAE) of 15.21, reflecting enhanced explanatory power and reduced cross-fold variability. The comparatively weaker performance of the Dual-Stream configuration suggests that structural separation without dynamic recombination fails to capture cross-stream complementarity under systematic missingness. These results underscore the importance of instance-level adaptivity in heterogeneous integration tasks. A benchmark comparison against XGBoost, a widely adopted gradient boosting framework for tabular prediction, further elucidates the role of availability-aware modeling. Although XGBoost demonstrates competitive aggregate performance, subset-level analysis reveals a statistically significant interaction between model architecture and digital data availability. In the subset of observations characterized by missing digital signals, the Fusion architecture significantly outperforms XGBoost (Fusion: 0.9182 vs. XGBoost: 0.8971 in \operatorname{R}^2), indicating enhanced robustness under high-uncertainty conditions. While tree-based models can route missing values through default splits, they do not explicitly learn stream-level reliability as a function of availability. In contrast, the gating mechanism in the Fusion architecture internalizes missingness structure, resulting in more stable and behaviorally interpretable predictions. Analysis of learned gating coefficients confirms that the model systematically increases reliance on the Sales-Only encoder under missingness conditions, empirically validating the theoretical premise that MNAR patterns can serve as adaptive reliability regulators within deep learning architectures. This finding bridges statistical missing-data theory with contemporary representation learning, demonstrating that structural absence can be leveraged as an informative supervisory signal. Collectively, this study advances predictive modeling in multi-source business environments along theoretical, methodological, and practical dimensions. Theoretically, it reconceptualizes MNAR missingness in digitally transformed ecosystems as a behaviorally meaningful construct rather than a preprocessing artifact. Methodologically, it introduces a scalable, dynamically gated Fusion Transformer architecture capable of expert specialization under heterogeneous reliability regimes. Practically, it demonstrates that organizations can integrate digitally enriched yet incomplete behavioral streams without compromising the stability of transactional foundations. By enabling models to adaptively determine when to rely on enriched auxiliary signals and when to default to stable historical baselines, the proposed framework supports resilient, interpretable, and high-accuracy forecasting in complex real-world operational contexts characterized by structured data asymmetry.
Benzer Tezler
- Sivas-Kangal yöresinde kırsal gelişme ve pazara açılma sürecindeki farklılaşmalar (bir örnek olay incelemesi)
Başlık çevirisi yok
FAHRİ ERCEM
Yüksek Lisans
Türkçe
1984
SosyolojiCumhuriyet ÜniversitesiSosyoloji Ana Bilim Dalı
DOÇ. DR. A. VAHAP SAĞ
- Sürekli rejim şartlarında nükleer yakıt çubuklarında tek ve iki boyutlu ısı iletiminin analizi
In steady-state conditions analyses of one and two dimensional conduction in nuclear fuel rods
İBRAHİM ATILGAN
Yüksek Lisans
Türkçe
1987
Makine MühendisliğiGazi ÜniversitesiMakine Mühendisliği Ana Bilim Dalı
DOÇ. DR. ERCAN ATAER
- Kutu kesitli kirişlerin eğilme, burulma ve bozulma etkilerine göre analizi
Başlık çevirisi yok
AHMET BUDAK
Yüksek Lisans
Türkçe
1987
İnşaat MühendisliğiKaradeniz Teknik Üniversitesiİnşaat Mühendisliği Ana Bilim Dalı
DOÇ. DR. İLHAN SUNGUR
- Süpersimetri
Başlık çevirisi yok
AYSUN HİLAL
Yüksek Lisans
Türkçe
1987
Fizik ve Fizik MühendisliğiÇukurova ÜniversitesiFizik Ana Bilim Dalı
PROF. DR. MEHMET KOCA