<p>Representation learning algorithms, as the cutting edge of modern AIs, has shown their ability to automatically solve complex tasks in diverse fields including computer vision, speech recognition, autonomous driving, biology. Unsurprisingly, representation learning applications in computer networking domains, such as network management, video streaming, traffic forecasting, are enjoying increasing interests in recent years. However, the success of representation learning algorithms is based on consistency between training and test data distribution, which can not be guaranteed in some scenario due to resource limitation, privacy or other infrastructure reasons. Caused by distribution shift in training and test data, representation learning algorithms have to apply tuned models into environments whose data distribution are solidly different from the model training. This issue is addressed as Out-Of-Distribution (OOD) Generalization, and is still an open topic in machine learning. In this dissertation, I present solutions for OOD cases found in cloud services which will be beneficial to improve user experience. First, I implement Infinity SGD which can extrapolate from light-load server log to predict server performance under heavy-load. Infinity SGD builds the bridge between light-load and heavy-load server status through modeling server status under different loads by an unified Continuous Time Markov Chain (CTMC) of same parameters. I show that Infinity SGD can perform extrapolations that no precedent works can do on real-world testbed and synthetic experiments. Next, I propose Veritas, a framework to answer what will be the user experience if a different ABR, a kind of video streaming data transfer algorithm, was used with the same server, client and connection status. Veritas strictly follows Structural Causal Model (SCM) which guarantees its power to answer what-if counterfactual and interventional questions for video streaming. I showcase that Veritas can accurately answer confounders for what-if questions on real-world emulations where on existing works can. Finally, I propose time-then-graph, a provable more expressive temporal graph neural network (TGNN) than precedent works. We empirically show that time-then-graph is a more efficient and accurate framework on forecasting traffic on network data which will serve as an essential input data for Infinity SGD. Besides, paralleling with this dissertation, I formalize Knowledge Graph (KG) as doubly exchangeable attributed graph. I propose a doubly exchangeable representation blueprint based on the formalization which enables a complex logical reasoning task with no precedent works. This work may also find potential traffic classification applications in networking field.</p>

In aviation, safety is a big concern. Knowing the position of an aircraft at all times is of high importance. Today most aircraft utilize Global Navigation Satellite Systems (GNSS) for localization and precision navigation because of the small position error which do not increase over time. However, recent research show that GNSS can easily be jammed or spoofed. An alternative navigation method is Visual Odometry (VO). VO is navigation through visual input and is a key-part in development of fully autonomous vehicles. This thesis investigates the Deep Learning-based Visual Odometry (DL-VO) for aircraft at altitudes over 100 m. DL-VO deployed at high altitude is almost none existing. Therefore, this thesis investigates the deployments of ground developed DL-VO in the aerial domain. DeepVO is a Frame-To-Frame optical flow estimation method which is trained supervised and end-to-end. The domain change, from ground to high altitude aerial, brought bigger issues and had larger impact on the performance than first though. The use of full 6 Degrees of Freedom (DoF) pose estimation increases the complexity and was much harder than 2D estimation (x, y, yaw). A good angle representation was of higher importance during training and testing in the aerial domain. Since in the aerial domain the full 3D rotation is not unique in all representations of the orientation and issues with Gimbal lock can occur. Results on simulated data show that the propose method fails to estimate 6 DoF poses. However, the reduced 2D estimations shows that a trajectory can be maintained even is drift is present. The result on real world dataset shows that it easier to recover scale at lower speeds and with a less down angled camera. The difference between simulated and non-simulated data has not been investigated to the extent that a fair assessment on how the method’s performance is affected by the data character. / Flygsäkerhet är av stor vikt inom flygindustin. Att som pilot alltid veta var planet befinner sig är av stor vikt. Global Navigation Satellite Systems (GNSS) är idag den mest använda metoden för lokalisering och precisionsnavigering då GNSS har liten felmarginal som inte förvärras över tid. Nyligen har forskare visat att GNSS kan lätt störas och alternativa lokaliseringsmetoder behövs. En av dem är Visual Odometry (VO). VO metoder försöker navigera sig i olika miljöer genom att estimera kamerors rörelser i sekvens av bilder. Det pågår mycket forsking på området då det är ett nyckelkoncept för autonoma fordon. Detta arbete undersöker användadet av Deep Learning-based Visual Odometry (DL-VO) för flygfarkoster på höjder över 100 m. Det är väldigt få som har testat DL-VO på annat än små drönare vilket skiljer från flygplan på högre höjd som stöter på andra problem där alla obejekt är väldigt små. Då forskingen på DL-VO för flygplan på högre höjd är minimal undersöker arbetet ett domän byte genom att ta en metod utveklad för markfordon och flytta den till flygdomänen. För att undersöka bytet av domän avändes en anpassad version av DeepVO nätverket. DeepVO använder sig av realtiv Frame-To-Frame optiskt flödes estimeringar och är tränad end-to-end enligt supervised learning metoden. Domän bytet, från mark till luft, medförde större problem än först trott och det ökade komplexiteten på problemet. Estimeringar med 6 frihetgrader är mer komplexa och en bra vinkel representation är av mycket större vikt. Minimering av vinklar under träningen skapade andra problem i flygdomänen än vad det gjorde på ursrungliga datasetet. Resultaten på simiulerad data visar att den framtagna metoden inte klarar estimeringar med 6 frihetgrader. Men om problemet reduceras så kan metoden estimaera 2D banor på en fixerad höjd i luften även om viss drift över tid existerar. Kameravinkeln och hastighet påverkar metodens förmåga att hålla en korrekt skala. Resultat på verklig data visar att det är lättare att uppnå korrekt skala vid lägre hastighet och mindre nervinklad kamera. Skillnaderna mellan simulerad och verklig data har inte undersökts i den utsträktning som behövs för att göra en korrekt slutsats om dess efftekter på resultatet.

<p>Temporal point processes (TPP) are mathematical approaches for modeling asynchronous event sequences by considering the temporal dependency of each event on past events and its instantaneous rate. Temporal point processes can model various problems, from earthquake aftershocks, trade orders, gang violence, and reported crime patterns, to network analysis, infectious disease transmissions, and virus spread forecasting. In each of these cases, the entity's behavior with the corresponding information is noted over time as an asynchronous event sequence, and the analysis is done using temporal point processes, which provides a means to define the generative mechanism of the sequence of events and ultimately predict events and investigate causality.</p> <p><br></p> <p>Among point processes, Hawkes process as a stochastic point process is able to model a wide range of contagious and self-exciting patterns. One of Hawkes process's well-known applications is predicting the evolution of viral processes on networks, which is an important problem in biology, the social sciences, and the study of the Internet. In existing works, mean-field analysis based upon degree distribution is used to predict viral spreading across networks of different types. However, it has been shown that degree distribution alone fails to predict the behavior of viruses on some real-world networks. Recent attempts have been made to use assortativity to address this shortcoming. This thesis illustrates how the evolution of such a viral process is sensitive to the underlying network's structure. </p> <p><br></p> <p>In Chapter 3, we show that adding assortativity does not fully explain the variance in the spread of viruses for a number of real-world networks. We propose using the graphlet frequency distribution combined with assortativity to explain variations in the evolution of viral processes across networks with identical degree distribution. Using a data-driven approach, by coupling predictive modeling with viral process simulation on real-world networks, we show that simple regression models based on graphlet frequency distribution can explain over 95\% of the variance in virality on networks with the same degree distribution but different network topologies. Our results highlight the importance of graphlets and identify a small collection of graphlets that may have the most significant influence over the viral processes on a network.</p> <p><br></p> <p>Due to the flexibility and expressiveness of deep learning techniques, several neural network-based approaches have recently shown promise for modeling point process intensities. However, there is a lack of research on the possible adversarial attacks and the robustness of such models regarding adversarial attacks and natural shocks to systems. Furthermore, while neural point processes may outperform simpler parametric models on in-sample tests, how these models perform when encountering adversarial examples or sharp non-stationary trends remains unknown. </p> <p><br></p> <p>In Chapter 4, we propose several white-box and black-box adversarial attacks against deep temporal point processes. Additionally, we investigate the transferability of white-box adversarial attacks against point processes modeled by deep neural networks, which are considered a more elevated risk. Extensive experiments confirm that neural point processes are vulnerable to adversarial attacks. Such a vulnerability is illustrated both in terms of predictive metrics and the effect of attacks on the underlying point process's parameters. Expressly, adversarial attacks successfully transform the temporal Hawkes process regime from sub-critical to into a super-critical and manipulate the modeled parameters that is considered a risk against parametric modeling approaches. Additionally, we evaluate the vulnerability and performance of these models in the presence of non-stationary abrupt changes, using the crimes and Covid-19 pandemic dataset as an example.</p> <p><br></p> <p> Considering the security vulnerability of deep-learning models, including deep temporal point processes, to adversarial attacks, it is essential to ensure the robustness of the deployed algorithms that is despite the success of deep learning techniques in modeling temporal point processes.</p> <p> </p> <p>In Chapter 5, we study the robustness of deep temporal point processes against several proposed adversarial attacks from the adversarial defense viewpoint. Specifically, we investigate the effectiveness of adversarial training using universal adversarial samples in improving the robustness of the deep point processes. Additionally, we propose a general point process domain-adopted (GPDA) regularization, which is strictly applicable to temporal point processes, to reduce the effect of adversarial attacks and acquire an empirically robust model. In this approach, unlike other computationally expensive approaches, there is no need for additional back-propagation in the training step, and no further network is required. Ultimately, we propose an adversarial detection framework that has been trained in the Generative Adversarial Network (GAN) manner and solely on clean training data. </p> <p><br></p> <p>Finally, in Chapter 6, we discuss implications of the research and future research directions.</p>

E-scooter riders have an increased crash risk compared to cyclists [1 ]. Hospital data finds increasing numbers of injured e-scooter riders, with head injuries as one of the most common injury types [2]. To decrease this high prevalence of head injuries, the use of e-scooter helmets could present a potential countermeasure [3]. Despite this, studies show a generally low rate of helmet use rates in countries without mandatory helmet use laws [4][5][6]. In countries with mandatory helmet use laws for e-scooter riders, helmet use rates are higher, but generally remain lower than bicycle use rates [7]. As the helmet use rate is a central factor for the safety of e-scooter riders in case of a crash and a key performance indicator in the European Commission's Road Safety Policy Framework 2021-2030 [8], efficient e-Scooter helmet use data collection methods are needed. However, currently, human observers are used to register e-scooter helmet use either in direct roadside observations or in indirect video-based observation, which is time-consuming and costly. In this study, a deep learning-based method for the automated detection of e-scooter helmet use in video data was developed and tested, with the aim to provide an efficient data collection tool for road safety researchers and practitioners.

Development of deep learning methods is dependent majorly on availability of large amounts of high quality data. To tackle the problem of data scarcity one of the workarounds is to generate synthetic data using deep learning methods. Especially, when dealing with trajectory data there are added challenges that come in to the picture such as high dependencies of the spatial and temporal component, geographical context sensitivity, privacy laws that protect an individual from being traced back to them based on their mobility patterns etc. This project is an attempt to overcome these challenges by exploring the capabilities of Generative Adversarial Networks (GANs) to generate synthetic trajectories which have characteristics close to the original trajectories. A naive model is designed as a baseline in comparison with a Long Short Term Memorys (LSTMs) based GAN. GANs are generally associated with image data and that is why Convolutional Neural Network (CNN) based GANs are very popular in recent studies. However, in this project an LSTM-based GAN was chosen to work with in order to explore its capabilities and strength of handling long-term dependencies sequential data well. The methods are evaluated using qualitative metrics of visually inspecting the trajectories on a real-world map as well as quantitative metrics by calculating the statistical distance between the underlying data distributions of the original and synthetic trajectories. Results indicate that the baseline method implemented performed better than the GAN model. The baseline model generated trajectories that had feasible spatial and temporal components, whereas the GAN model was able to learn the spatial component of the data well and not the temporal component. Conditional map information could be added as part of training the networks and this can be a research question for future work. / Utveckling av metoder för djupinlärning är till stor del beroende av tillgången på stora mängder data av hög kvalitet. För att ta itu med problemet med databrist är en av lösningarna att generera syntetisk data med hjälp av djupinlärning. Speciellt när man hanterar bana data finns det ytterligare utmaningar som kommer in i bilden såsom starka beroenden av den rumsliga och tidsmässiga komponenten, geografiska känsliga sammanhang, samt integritetslagar som skyddar en individ från att spåras tillbaka till dem baserat på deras mobilitetsmönster etc. Detta projekt är ett försök att överkomma dessa utmaningar genom att utforska kapaciteten hos generativa motståndsnätverk (GAN) för att generera syntetiska banor som har egenskaper nära de ursprungliga banorna. En naiv modell är utformad som en baslinje i jämförelse med en LSTM-baserad GAN. GAN:er är i allmänhet förknippade med bilddata och det är därför som CNN-baserade GAN:er är mycket populära i nya studier. I det här projektet valdes dock en LSTM-baserad GAN att arbeta med för att utforska dess förmåga och styrka att hantera långsiktiga beroenden och sekventiella data på ett bra sätt. Metoderna utvärderas med hjälp av kvalitativa mått för att visuellt inspektera banorna på en verklig världskarta samt kvantitativa mått genom att beräkna det statistiska avståndet mellan de underliggande datafördelningarna för de ursprungliga och syntetiska banorna. Resultaten indikerar att den implementerade baslinjemetoden fungerade bättre än GAN-modellen. Baslinjemodellen genererade banor som hade genomförbara rumsliga och tidsmässiga komponenter, medan GAN-modellen kunde lära sig den rumsliga komponenten av data väl men inte den tidsmässiga komponenten. Villkorskarta skulle kunna läggas till som en del av träningen av nätverken och detta kan vara en forskningsfråga för framtida arbete.

<p>Deep Neural Networks (DNNs) are constantly evolving, enabling the power of deep learning to be applied to an ever-growing range of applications, such as Natural Language Processing (NLP), recommendation systems, graph processing, etc. However, these emerging neural workloads present large computational demands for both training and inference. In this dissertation, we propose optimizations that take advantage of the unique characteristics of different emerging workloads to simultaneously improve accuracy and computational efficiency.</p> <p><br></p> <p>First, we consider Language Models (LMs) used in NLP. We observe that the design process of LMs (pre-train a foundation model, and subsequently fine-tune it for different downstream tasks) leads to models that are highly over-parameterized for the downstream tasks. We propose AxFormer, a systematic framework that applies accuracy-driven approximations to create accurate and efficient LMs for a given downstream task. AxFormer eliminates task-irrelevant knowledge, and helps the model focus only on the relevant parts of the input.</p> <p><br></p> <p>Second, we find that during fine-tuning of LMs, the presence of variable-length input sequences necessitates the use of padding tokens when batching sequences, leading to ineffectual computations. It is also well known that LMs over-fit to the small task-specific training datasets used during fine-tuning, despite the use of known regularization techniques. Based on these insights, we present TokenDrop + BucketSampler, a framework that synergistically combines a new regularizer that drops a random subset of insignificant words in each sequence in every epoch, and a length-aware batching method to simultaneously reduce padding and address the overfitting issue.</p> <p><br></p> <p>Next, we address the computational challenges of Transformers used for processing inputs of several important modalities, such as text, images, audio and videos. We present Input Compression with Positional Consistency (ICPC), a new data augmentation method that applies varying levels of compression to each training sample in every epoch, thereby simultaneously reducing over-fitting and improving training efficiency. ICPC also enables efficient variable-effort inference, where easy samples can be inferred at high compression levels, and vice-versa.</p> <p><br></p> <p>Finally, we focus on optimizing Graph Neural Networks (GNNs), which are commonly used for learning on non-Euclidean data. Few-shot learning with GNNs is an important challenge, since real-world graphical data is often sparsely labeled. Self-training, wherein the GNN is trained in stages by augmenting the training data with a subset of the unlabeled data and their pseudolabels, has emerged as a promising approach. However, self-training significantly increases the computational demands of training. We propose FASTRAIN-GNN, a framework for efficient and accurate self-training of GNNs with few labeled nodes. FASTRAIN-GNN optimizes the GNN architecture, training data, training parameters, and the graph topology during self-training.</p> <p><br></p> <p>At inference time, we find that ensemble GNNs are significantly more accurate and robust than single-model GNNs, but suffer from high latency and storage requirements. To address this challenge, we propose GNN Ensembles through Error Node Isolation (GEENI). The key concept in GEENI is to identify nodes that are likely to be incorrectly classified (error nodes) and suppress their outgoing messages, leading to simultaneous accuracy and efficiency improvements. </p> <p><br></p>

Cell culture is a fundamental procedure in many laboratories and precedes much research performed under the microscope. Despite the significance of this procedural stage, the monitoring of cells throughout growth is impossible due to the absence of equipment and methodological approaches. This thesis presents a low-cost, power-effective and versatile microscope with small enough dimensions to operate inside an incubator. Besides image acquisition, the microscope comprises other functions such as a data pipeline, implemented to save the images on the user’s computer via a server whilst also offering storage of the images on an integrated micro SD-card. Furthermore, a machine learning algorithm with a human-in-the-loop approach has been trained to segment the acquired images for cell proliferation and cell apoptosis tracking, and yielded promising results with an accuracy of 94%. For comparison, conventional segmentation techniques using operations such as the watershed function were deployed.The microscope described is versatile in operation as it offers the user to utilise one or more functions, depending on the purpose of the imaging. / Cellodling är en grundläggande process i många laboratiorium och föregår forskning som utförs under mikroskop. Trots inkubationens betydelse har övervakning av celler i detta skede inte varit möjlig på grund utav avsaknaden av relevant utrustning och metodologiska tillvägagångsätt. I denna examensuppsatts på avancerad nivå presenteras ett lågkostnads-, energieffektivt och versatilt mikroskop av centimeterstora dimensioner anpassat för användning i en inkubator. Förutom bildtagningsmekanismer erbjuder mikroskopet olika funktioner som till exempel ett integrerat dataflöde som möjliggör sparande av bilder på användarens dator via en server samtidigt som den erbjuder sparande av bilder på ett integrerat minneskort.Utöver detta har en human-in-the-loop maskininlärningsalgoritm för segmentation av celler implementerats i syfte att övervaka cellernas celldelning och celldöd. Denna algoritm påvisade goda resultat med en nogrannhet på 94%. I jämförelsesyfte har även en traditionell watershed-baserad cellsegmenteringsteknik utvecklats.Mikroskopet kan kallas versatilt då det tillåter användaren att anpassa dataflödet och välja vilka funktioner denne vill nyttja, allt utefter bildtagningens ändamål.

In the brain, neuronal cells located in different functional regions communicate through a dense structural network of axons known as the white matter (WM) tissue. Bundles of axons that share similar pathways characterize the WM anatomy, which can be investigated in-vivo thanks to the recent advances of magnetic resonance (MR) techniques. Diffusion MR imaging combined with tractography pipelines allows for a virtual reconstruction of the whole WM anatomy of in-vivo brains, namely the tractogram. It consists of millions of WM fibers as 3D polylines, each approximating thousands of axons. From the analysis of a tractogram, neuroanatomists can characterize well-known white matter structures and detect anatomically non-plausible fibers, which are artifacts of the tractography and often constitute a large portion of it. The accurate characterization of tractograms is pivotal for several clinical and neuroscientific applications. However, such characterization is a complex and time-consuming process that is difficult to be automatized as it requires properly encoding well-known anatomical priors. In this thesis, we propose to investigate the encoding of anatomical priors with a supervised deep learning framework. The ultimate goal is to reduce the presence of artifactual fibers to enable a more accurate automatic process of WM characterization. We devise the problem by distinguishing between volumetric and non-volumetric representations of white matter structures. In the first case, we learn the segmentation of the WM regions that represent relevant anatomical waypoints not yet classified by WM atlases. We investigate using Convolutional Neural Networks (CNNs) to exploit the volumetric representation of such priors. In the second case, the goal is to learn from the 3D polyline representation of fibers where the typical CNN models are not suitable. We introduce the novelty of using Geometric Deep Learning (GDL) models designed to process data having an irregular representation. The working assumption is that the geometrical properties of fibers are informative for the detection of tractogram artifacts. As a first contribution, we present StemSeg that extends the use of CNNs to detect the WM portion representing the waypoints of all the fibers for a specific bundle. This anatomical landmark, called stem, can be critical for extracting that bundle. We provide the results of an empirical analysis focused on the Inferior Fronto-Occipital Fasciculus (IFOF). The effective segmentation of the stem improves the final segmentation of the IFOF, outperforming with a significant gap the reference state of the art. As a second and major contribution, we present Verifyber, a supervised tractogram filtering approach based on GDL, distinguishing between anatomically plausible and non-plausible fibers. The proposed model is designed to learn anatomical features directly from the fiber represented as a 3D points sequence. The extended empirical analysis on healthy and clinical subjects reveals multiple benefits of Verifyber: high filtering accuracy, low inference time, flexibility to different plausibility definitions, and good generalization. Overall, this thesis constitutes a step toward characterizing white matter using deep learning. It provides effective ways of encoding anatomical priors and an original deep learning model designed for fiber.

The aim of this work is to compare methods that predict whether an image has suggestive content, such as pornographic images and erotic fashion. Using binary classification, this work contributes to an internet environment where these images are not seen out of context. It is, therefore, necessary for user experience improvement purposes, such as child protection, publishers not having their campaign associated with inappropriate content, and companies improving their brand safety. For this study, a data set with more than 500k images was created to test the Convolutional Neural Networks (CNN) models: NSFW model, ResNet, EfficientNet, BiT, NudeNet and Yahoo Model. The image classification model EfficientNet-B7 and Big Transfer (BiT) presented the best results with over 91% samples correctly classified on the test set, with precision and recall around 0.7. Model prediction was further investigated using Local Interpretable Model-agnostic Explanation (LIME), a model explainability technique, and concluded that the model uses coherent regions of the thumbnail according to a human perspective such as legs, abdominal, and chest to classify images as unsafe. / Syftet med detta arbete är att jämföra metoder som förutsäger om en bild har suggestivt innehåll, såsom pornografiska bilder och erotiskt mode. Med binär klassificering bidrar detta arbete till en internetmiljö där dessa bilder inte ses ur sitt sammanhang. Det är därför nödvändigt för att förbättra användarupplevelsen, till exempel barnskydd, utgivare som inte har sina kampanjer kopplade till olämpligt innehåll och företag som förbättrar deras varumärkessäkerhet. För denna studie skapades en datamängd med mer än 500 000 bilder för att testa Convolutional Neural Networks (CNN) modeller: NSFW-modell, ResNet, EfficientNet, BiT, NudeNet och Yahoo-modell. Bild klassificerings modellen EfficientNet-B7 och Big Transfer (BiT) presenterade de bästa resultaten med över 91%prover korrekt klassificerade på testuppsättningen, med precision och återkallelse runt 0,7. Modell förutsägelse undersöktes ytterligare med hjälp av Local Interpretable Model-agnostic Explanation (LIME), en modell förklarbarhetsteknik, och drog slutsatsen att modellen använder sammanhängande regioner i miniatyren enligt ett mänskligt perspektiv såsom ben, buk och bröst för att klassificera bilder som osäkra.

[pt] O problema de Districting-and-routing é um problema estratégico no qual porções geográficas devem ser agregadas em regiões de entrega, e cada região de entrega possui um custo de roteamento estimado. Seu objetivo é de minimizar esses custos, além de garantir a divisão da região em distritos. A simulação para obter uma boa aproximação é muito custosa computacionalmente, enquanto mecanismos como buscas locais exigem que esse cálculo seja feito de forma muito eficiente, tornando essa estratégia de aproximação inviável para uma solução metaheurística. Grande parte das soluções existentes para esse problema utilizam de formulas de aproximação contínua para mensurar os custos de roteamento, funções essas que são rápidas de serem calculadas porém cometem erros significativos. Em contraste, propomos uma Rede Neural em Grafo (Graph Neural Network - GNN) que é usada como oráculo por um algoritmo de otimização. Nossos experimentos computacionais executados com dados de cidades do Reino Unido mostram que a GNN é capaz de produzir previsões de custos mais precisas em tempo computacional aceitável. O uso desse estimator na busca local impacta positivamente a qualidade das soluções, levando a uma economia de 10,35 por cento no custo de entrega estimado em relação a função Beardwood, que é comumente usada nesse cenários, e ganhos similares em comparação com outros métodos de aproximação. / [en] The districting-and-routing problem is a strategic problem in which basic geographical units (e.g., zip codes) should be aggregated into delivery regions, and each delivery region is characterized by a routing cost estimated over an extended planning horizon. The objective is to minimize the expected routing costs while ensuring regional separability through the definition of the districts. Repeatedly simulating routing costs on a set of scenarios while searching for good districts can be computationally intensive, so existing solution approaches for this problem rely on approximation functions. In contrast, we propose to rely on a graph neural network (GNN) trained on a set of demand scenarios, which is then used within an optimization approach to infer routing costs while solving the districting problem. Our computational experiments on various metropolitan areas show that the GNN produces accurate cost predictions. Moreover, using this better estimator during the search positively impacts the quality of the districting solutions and leads to 10.35 percent delivery-cost savings over the commonly-used Beardwood estimator and similar gains compared to other approximation methods.