Umeå University's logo

This thesis was conducted within the project Choking Under Pressure, a longitudinal study following Swedish Grade 4 students during four semesters. The main purpose of this thesis was to develop and validate a math anxiety rating scale for Swedish-speaking primary school students in a Nordic context, and to explore the underlying mechanisms in the math anxiety–math performance relationship. This thesis consists of four studies, where Studies I and II adopt a measurement perspective. Study III provides a review of the role of working memory in the math anxiety–math performance relationship. Study IV extends Study III by empirically examining working memory and self-concept as mediators.

Study I assessed the validity and reliability of the Swedish MARS-E. Test-retest reliability and internal consistency demonstrated longitudinal stability and strong item agreement. Dimensionality analysis supported a one-factor structure, with evidence of both gender and longitudinal invariance. Significant gender differences in math anxiety were found, increasing across time points.

Study II extended Study I by combining data from two Finnish projects–one Finnish and one Finland-Swedish–with the Swedish sample. All measurements were conducted in Spring 2023 on Grade 4 students using the same instruments. A one-factor model was found for both Swedish-speaking groups, whereas the Finnish version of the MARS-E supported a two-factor model.

Study III was a literature review employing meta-analysis to examine the relationship between working memory and math anxiety, and the mediating role of working memory in the math anxiety–performance link. Results from 57 studies showed a significant negative correlation, and data from 8 studies indicated a partial mediation effect.

Study IV empirically examined how working memory and math self-concept influence the math anxiety–math performance link. Using structural equation models across two longitudinal waves, as well as multiple mediation, the results showed that both working memory and self-concept in math are important mediators in this relationship.The main contributions of this thesis are the development and validation of the Swedish MARS-E, as well as a deeper understanding of the mediating roles of working memory and math self-concept in the math anxiety–math performance relationship. These findings have implications for both research and educational practice, refining theoretical models of math anxiety and its cognitive and motivational correlates, and providing a foundation for interventions targeting math anxiety.

The purpose of this dissertation is to critically evaluate and examine the concept of wilderness romance [vildmarksromantik], its various possibilities, meanings, and uses in a Northern Swedish literary context. Theories on modernity and ecocritisism are important starting points for the analysis. The category “wilderness romance” has been used to classify early Northern Swedish literature from the turn of the last century. However, the term "wilderness romance" has often carried a pejorative meaning, depicting Northern Sweden in a romantic light with a focus on beautiful nature — far from the social problems that characterized the region in the wake of modernity and the industrialization of forestry, hydropower, and mining.

The thesis starts with a reception study of the author Pelle Molin’s book Ådalens poesi (1897) where the origins of the meaning behind the concept of “wilderness romance” is traced. The study proceeds to examine how “wilderness romance” has been used and established as a pejorative concept in both press and in literary history. The study finds that the negative response to wilderness romance in the historical works draws on the romanticizing view of Ådalens poesi and Pelle Molin present in the reception rather than readings of Molin or some of his allegedly many epigones.

The thesis continues to analyze a few of the literary works that has been categorized as “wilderness romance”, drawing from theories on modernity and ecocriticism. The analysis shows that none of the works are silent on the matter of industrialization and modernization of Northern Sweden around the turn of the century. Rather, a criticism based in the magical and otherworldly emerges, highlighting the importance of human immanence in relation to nature.

Further, the thesis moves on to analyse a few contemporary northern Swedish novels to track in what way traces of “wilderness romance” can be found in contemporary works. The study finds that also the contemporary works utilizes some of the same motifs to carry the criticism of civilization, such as the myths of trolls, the deep and unknown forest and the problematization of the human-animal relationship.

In conclusion, the thesis finds that the previous understanding of the category “wilderness romance” to be simplified regarding the critical potential of the works of literature categorized as such. There is, in these works, criticism of modernity and exploitation of nature, but also a story about grief and sorrow of the enchanted world lost through industrialization.

This dissertation examines the dynamics of Black Open Access, a pirate-driven phenomenon, addressing inequities in academic publishing through shadow libraries and text piracy. Through a methodological patchwork combining netnography, computational methods, and text analysis, this dissertation investigates how these phenomena operate at the intersection of formal and informal media economies. The results show how shadow libraries like Sci-Hub, Library Genesis, and Z-library are more than simple piracy platforms, and should be viewed as robust ecosystems with their own technical infrastructure, community norms, and justificatory frameworks. The findings demonstrate that Black Open Access solutions persist through ”Pirate LOCKSS”. A decentralized preservation strategy utilizing multiple domain copies and established internet platforms as intermediaries. These communities develop complex legitimization mechanisms, from gamified user engagement systems to quasi-legal frameworks that mimic traditional academic institutions. Users justify their participation through multifaceted moral arguments about knowledge democratization and academic freedom, balanced with practical necessities driven by institutional constraints.

The dissertation shows that rather than operating in mere opposition to formal academic publishing, shadow libraries function as parallel systems that both challenge and complement traditional knowledge distribution. This creates a paradox of legitimacy wherein Black Open Access initiatives simultaneously reject copyright frameworks while reproducing many norms and practices of the formal academic system.  By analyzing these dynamics, this dissertation contributes to understanding how informal media economies function in academic contexts and demonstrates how shadow libraries have become embedded in scholarly workflows, creating an alternative infrastructure for knowledge dissemination that responds to structural failures in academic publishing while raising important questions about the future of scholarly communication.

Background: In Tanzania, 13% of families have a member with a disability, presenting unique challenges for entire household. Despite its complexity, disability is often overlooked in development and research agendas. However, disability has gained significant attention over the past decade, particularly with its integration in the 2030 sustainable development goals that were adopted by all United Nations member states in 2015. These goals are intended to promote inclusion and equal opportunities for all, including young people with disabilities. Nevertheless, young people with physical disabilities face considerable barriers to participation in society, driven by negative norms, discrimination, and stigma. This population remains underrepresented in research, and there is a significant gap in public health knowledge, particularly in resource-poor settings. This thesis aims to describe and explore the needs and challenges associated with being young and having a physical disability, drawing on the perspectives of children, adolescents, and young adults with disabilities – along with their families and carers – in northeastern Tanzania.

Methods: The research was conducted in the Kilimanjaro region of northeastern Tanzania, using both qualitative and quantitative methods. A quantitative survey was completed by the primary carers (aged 18 years or older) of 212 children and adolescents with physical disabilities, aged between 2 and 18 years. The data collected during the survey were analysed in Studies I and III. The SurveyCTO software was used to collect data electronically via a questionnaire. The questionnaire was developed based on the International Classification of Functioning, Disability, and Health (ICF) framework. Statistical analysis (Studies I and III) was performed using IBM SPSS version 28. For the qualitative aspect of the research, Study II employed constructivist grounded theory, while Study IV analysed the experiences of adolescents and young adults regarding inclusion using reflexive thematic analysis.

Results: Children and adolescents with physical disabilities have significant difficulties with walking, self-care, and participating in social activities, household chores, and schooling. Furthermore, they were reported needing assistive devices, therapeutic exercises, and nutritional support. Regarding essential services, most families lived near health facilities that did not provide rehabilitation services. Additionally, a quarter of the children and adolescents with physical disabilities had never received any rehabilitation service. More than two thirds did not have health insurance. The carers faced psychological challenges, with over half of the 212 carers surveyed reporting poor psychological wellbeing, and 42% and 38% describing symptoms of anxiety and depression, respectively. Support from extended family members was linked to better psychological wellbeing and a lower likelihood of experiencing anxiety and depression. Support systems, particularly via social networks, were crucial in helping families meet basic needs and ensure the safety of vtheir children. The adolescents and young adults, however, reported feeling marginalised, facing significant barriers to full societal integration, and experiencing an uncertain future with little hope.

Conclusion: This thesis found significant activity limitations among children and adolescents with physical disabilities in areas such as walking, self-care, communication, and participation in social activities, household chores, and schooling. Access to essential services, particularly healthcare, is inadequate, or in many cases impossible to access. Adolescents and young adults face both present and future barriers to full inclusion in society, and families are constantly adapting to meet the challenging needs of their children, with carers experiencing a considerable psychological burden. To address these challenges, it is crucial for the Ministry of Health to strengthen rehabilitation services within primary healthcare settings. Health-insurance reforms are also necessary to ensure the societal inclusion of young people with disabilities. Public-awareness campaigns that highlight the potential of people with disabilities and promote inclusion are crucial for supporting social networks and challenging negative societal norms. 

The green transition is the great undertaking of our time, and it will require significant ingenuity and change in all areas of society. Most urgently, perhaps, regarding energy, where the demand for transport, electricity and heat must be met by renewables instead of fossil fuels. Solar thermal is one alternative with the potential to contribute substantially to sustainable heat production. To realize this potential, the availability of competitive, sustainable and cost effective optical coatings for solar collectors is a prerequisite. The coatings used today are primarily produced with expensive vacuum-based deposition techniques, transferring a hampering cost to the collectors, which impede the deployment of solar thermal as an energy source. Herein, we show that by leveraging scalable deposition techniques, with elaborate material choices and innovative nanoscale designs, it is possible to produce sustainable coatings that are highly competitive with regards to cost and performance.Using a scalable aerosol-based deposition technique, an antireflective mesoporous silica coating, commonly implemented in advanced solar technologies, is produced with an ordered hexagonal pore structure. The attention to optical thickness and pore structure facilitates a superior performance and an increased durability, making it especially suitable for arid climates. Moreover, we present several methods to achieve solar selectivity for the receiver. We leverage the large potential window of a deep eutectic solvent to facilitate electrodeposition of a texture-based cobalt-chromium coating, making an otherwise unsustainable technique viable today. High selectivity is also achieved by manipulating interference effects in coatings produced through precise control of thermal annealing of steel and ultrasonic spray coating of carbon nanotube composites. Such optical effects are only achieved for selective coatings deposited with more advanced and expensive techniques.Science is an iterative process of small incremental advances, often seemingly insignificant in the moment, which over time accumulate to surprisingly quick change. Here we present examples of sustainable, scalable, durable and cost competitive antireflective and solar selective coatings, thereby hopefully contributing to an accelerated implementation of solar thermal technologies.

Machine learning models are inherently shaped by the data used to train them. Understanding the relationship between datasets and the models they generate is essential for tasks such as model selection, privacy metrics, and robustness evaluation. This thesis presents a rigorous mathematical framework for comparing machine learning models and algorithms by formalizing the interaction between two fundamental spaces: the database space, which captures possible datasets, and the model space, which contains the models or hypotheses derived from those datasets. A central motivation stems from the observation that different datasets can lead to the same or highly similar models. Such recurrent models—which arise frequently across diverse data sources—are particularly significant in privacy-sensitive applications. Their recurrence suggests reduced dependence on any specific data point or subgroup, thus offering inherent privacy and generalization benefits. By quantifying the relationship between models and their generating data, this work enables principled evaluation of a model’s robustness and disclosure risk.

To formalize relationships between the two spaces, the thesis develops a family of probabilistic metric space constructions tailored to different aspects of the data–model interaction. The first contribution models database evolution as a Markov process and defines probabilistic distances between models based on the likelihood of transitioning between their generating datasets. The second contribution introduces F-space, a framework based on fuzzy measures that captures richer structural properties of the data—such as redundancy, synergy, and overlap among subsets. Building on this, the third contribution applies the F-space theory in practical machine learning scenarios. It demonstrates how fuzzy measures can be used to compare different linear regression algorithms trained over structured subsets of real datasets. The final contribution further generalizes the framework through Generalized F-spaces, where the model space itself is endowed with probabilistic structure—allowing uncertainty in both the datasets and the model outputs to be captured simultaneously.

Together, these constructions offer a principled alternative to traditional model comparison metrics. Rather than relying solely on pointwise loss or accuracy, the proposed framework incorporates the diversity, dynamics, and internal structure of the data that underlies each model—enabling more robust and privacy-aware assessments.

Despite significant advancements in prevention and treatment, atherosclerotic cardiovascular disease remains a leading cause of mortality and morbidity. Atherosclerosis develops from the accumulation of lipoprotein debris in arterial walls, resulting in plaque buildup that causes arterial narrowing, thickening, or softening and may ultimately trigger thrombosis. Current therapies effectively lower low-density lipoprotein (LDL) levels while insufficiently addressing other atherogenic lipids like very-low-density lipoproteins (VLDL) and chylomicron remnants. Furthermore, the optimal timing for initiating lipid-lowering interventions is debated. Conventional cardiovascular prevention strategies, which base treatment decisions on ten-year risk calculations, may underestimate the cumulative impact of lifelong lipid exposure.

This thesis uses human genetics to explore the lifelong impact of inhibiting specific lipid-lowering drug candidate targets. We examine two key approaches in lipoprotein lowering: activating the rate-limiting enzyme in intravascular triglyceride hydrolysis, lipoprotein lipase (LPL), focusing on its activation through inhibiting the angiopoietin-like (ANGPTL) protein family of regulators; and the reverse cholesterol transport system, reevaluating cholesteryl ester transfer protein (CETP) as a drug target.

Through genetic association studies, Mendelian randomization, genetic mimicry analyses, and meta-analyses of clinical trials, we demonstrate that targeting these proteins may offer protection against atherosclerotic cardiovascular disease. Our findings support the ongoing clinical development of ANGPTL3, ANGPTL4, and CETP inhibitors for cardiovascular prevention while emphasizing the value of human genetics in drug discovery. Lastly, this work improves our understanding of lipid management throughout the lifespan and highlights the potential benefits of early intervention.   

Paper [I]: This paper studies the effect of Green Public Procurement (GPP) on competition, bids, and winning bids under two different regulation periods where the latter include more explicitly expressed GPP ambitions. Based on detailed data from Swedish internal cleaning service procurements, our results imply that environmental considerations might not influence the bids as required for GPP to be considered an effective environmental policy instrument. Over time, lower degree of competition and increased bids are found. This phenomenon can be attributed, at least in part, to regulatory influences, signifying an escalating complexity in the process of submitting bids.

Paper [II]: We combine theoretical and empirical analysis to investigate impacts of contractual requirements in procurement auctions with endogenous entry. Our analysis shows that contractual requirements impact expected payoffs, influencing the equilibrium number of bidders and their bids under zero-profit conditions. Specifically, we find that, in equilibrium, more contractual requirements increase competition by raising bidders’ expected payoffs, while increased entry costs reduce competition. Overall, a rise in contractual requirements and entry costs results in higher equilibrium bids. Additionally, we highlight the importance of enforce ability in shaping bidding behavior. Empirical evidence from public cleaning services procurement in Sweden corroborates the implications of our model.

Paper [III]: This paper investigates post-litigation impacts on bidder participation and composition within public procurement, utilizing Swedish data spanning from 2012 to 2018. The findings indicate an average decrease in bidder participation following a litigation process of approximately 8%, particularly in construction procurements, alongside a decline in the proportion of SME bidders and local firm participation. Furthermore, bidder participation, bidder composition, and procurement design significantly influence the probability of litigation.

Paper [IV]: This study examines the impact of local market concentration on the participation and success of small and medium enterprises (SMEs) in Swedish municipal cleaning service procurement auctions. A 10 percentage point reduction in the joint market share of the four largest firms (CR4), while maintaining a constant Herfindahl-Hirschman Index (HHI), results in a 7.5% increase in SME participation and raises the likelihood of an SME winning by 2.4%. Furthermore, the 2014 revisions to the EU public procurement directives mitigated the adverse effects of market concentration. However, despite the increase in participation, there is no evidence that the success rates of SMEs improved following the implementation of the revised EU directives. 

As technology becomes an integral part of everyday life for many, it also becomes a tool to facilitate harm and injustice towards communities who face marginalization.

To understand and address challenges of systemic oppression, the field of human–computer interaction (HCI) has moved towards framing larger questions of injustice and systemic issues. Through the growing area of social justice in HCI, researchers are collectively working to frame more just futures for communities that have experienced marginalization. 

However, the influence of systemic oppression on socio-technical systems can create hard-to-predict outcomes that ultimately lead to reproducing harmful practices towards marginalized communities. 

Through the included papers in this dissertation, on a range of topics from cultural heritage for Sámi communities, ageism reproduced through technology implementation, socio-political awareness to mitigate technology-facilitated sex trafficking, supporting women with experience of sexual violence to paths of justice, building a cohesive and fluid understanding of social justice in HCI, and critiquing interpersonal safety technologies, I present and discuss how the influence of systemic oppression creates hard-to-predict outcomes in the socio-technical systems we create. I show how these outcomes can lead to reproducing harmful practices towards marginalized communities and why it is important to actively work against this. 

To address the outcomes and harmful practices that I refer to as the “harm-reproduction loop,” I propose a theoretical lens called systemic sensitivity. Systemic sensitivity is a lens to support researchers in understanding the loop, addressing hard-to-predict outcomes and mitigating harmful practices. In doing so, we can more actively work against systemic oppression and work towards socio-political change and long-term contributions towards marginalized communities. 

This dissertation is a compilation dissertation, which means it is based on six papers in total. These papers are the foundation of the work produced, and all have been conducted within social justice in HCI.

This thesis analyses the development and transformation of Sweden’s contraceptive market between 1910 and 1979, a period marked by restrictive regulations (1910–1970) and intense moral contestation. The aim is to generate new knowledge about the historical evolution of the contraceptive market and to shed light on how a controversial market, characterised by strong moral and political stances, can become socially accepted over time. The thesis focuses on the role that companies played in this transformation, how they navigated the rules and norms surrounding the market, and how the production and sales methods of contraceptives have changed throughout the period.

The study combines two analytical frameworks: The Architecture of Markets – which examines how markets are created, shaped, and change over time – and Hostile Worlds – a framework for understanding how boundaries between commerce and intimacy generate controversy. In doing so, the thesis analyses the mechanisms through which the controversial market both expanded and gained legitimacy. Drawing on scholarship from business history, consumer history, and the history of sexuality, the thesis explores the interplay between shifting sexual norms, state regulations, and commercial strategies in shaping market dynamics. By analysing sexuality and gender norms as constituent elements in the architecture of the contraceptive market, the thesis also provides insights into the bidirectional relationship within sexual markets. It demonstrates how these markets are shaped by, and in turn help shape, broader cultural and moral values.

The study addresses three key questions: What factors enabled or constrained the development of the contraceptive market between 1910 and 1979? How and why did the production and retail of mass-produced contraceptives change during this period? How did controversies arise within and around the market, and how did businesses navigate these tensions? While previous studies on the business history of contraception have primarily focused on countries with large-scale contraceptive manufacturers, this study shifts attention to Sweden, where the market expanded largely through the activities of retailers and importers rather than domestic producers. 

Based on a wide range of archival sources, including business records, advertisements, government reports, parliamentary and legal documents, letters, newspapers, descriptive statistics, and photographs, the findings demonstrate that Swedish businesses employed both provocative and legitimising strategies to navigate legal and social market restrictions. Mail-order services and vending machines, for example, allowed companies to balance discretion with accessibility. Marketing strategies deliberately framed contraceptives as medical necessities or as products tied to sexual well-being, adapting to prevailing moral discourses while also challenging norms on sexuality, gender, and consumption.

This thesis argues that the expansion of the contraceptive market was not exclusively driven by governmental reforms, shifting sexual norms, or political activism; it was also significantly influenced by business strategies that pushed the boundaries of what was perceived as acceptable sexuality and legitimate business. It further demonstrates how market controversies were not only shaped by the symbolic values attached to commodities but also by notions surrounding distribution methods, sales practices, and consumer demand. By situating Sweden’s contraceptive market at the intersection of commerce and sexuality, this study contributes to broader discussions on how controversial markets gain legitimacy over time and how businesses contribute to shaping societal attitudes towards consumption and morality.

In the rapidly evolving era of Artificial Intelligence (AI), privacy-preserving techniques have become increasingly important, particularly in areas such as deep learning (DL). Deep learning models are inherently data-intensive and often rely on large volumes of sensitive personal information to achieve high performance. The increasing usage of data has raised critical concerns around data privacy, user control, and regulatory compliance. As data-driven systems grow in complexity and scale, safeguarding individual privacy while maintaining model utility has become a central challenge in the field of deep learning. Traditional privacy-preserving frameworks focus either on protecting privacy at the database level (e.g., k-anonymity) or during the inference/output from the model (e.g., differential privacy), each introducing trade-offs between privacy and utility. While these approaches have contributed significantly to mitigating privacy risks, they also face practical limitations, such as vulnerability to inference attacks or degradation in model performance due to the added noise. 

In this thesis, we take a different approach by focusing on anonymous models (i.e., models that can be generated by a set of different datasets) in model space with integral privacy. Anonymous models create ambiguity to the  intruder by ensuring that a trained model could plausibly have originated from various datasets, at the same time it gives the flexibility of choosing models which do not cause much utility loss. Since exhaustively exploring the model space to find recurring models is computationally intensive, we introduce a relaxed variant called ∆-Integral Privacy, where two models are considered recurring if they are within a bounded ∆ distance. Using this notion, we present practical frameworks for generating integrally private models for both machine learning and deep learning settings. We also provide probabilistic guarantees, demonstrating that under similar training environments, models tend to recur with high probability when optimized using mean samplers (e.g., SGD, Adam). These recurring models can be further utilized as an ensemble of private model that can estimate prediction uncertainty, which can be used for privacy-preserving concept drift detection in streaming data. We further extend our investigation to distributed settings, particularly Federated Learning (FL), where a central server only aggregates client-side model updates without accessing raw data. With strong empirical evidences supported by theoretical guarantees, we can claim that our frameworks with integral privacy are robust alternatives to conventional privacy-preserving methods.

In addition to generating anonymous models, this thesis also focus on developing approaches which enable users to remove their data and their influence from a machine learning model (machine unlearning) under their right-to-be-forgotten. As the demand for data removal and compliance with right-to-be-forgotten regulations intensifies, the need for efficient, auditable, and realistic unlearning mechanisms becomes increasingly urgent. Beyond regulatory compliance, machine unlearning plays a vital role in removing copyrighted content as well as mitigating security risks. In federated learning, when the clients share same feature space and participate in a huge number, we argue that if the server employs integrally private aggregation mechanism then it can plausibly deny client participation in training the global model up to certain extent. Hence, reducing the computational requirements for frequent unlearning requests. However, when there are limited number of clients with complimentary features, the server must employ unlearning mechanisms to deal with the model compression and high communication cost. This thesis shows that machine unlearning can be made efficient, effective and auditable, even in complex, distributed and generative environments. Our work spans across multiple dimensions, including privacy, auditability, computational efficiency, and adaptability across diverse data modalities and model architectures. 

Rigidity theory is the mathematical study of rigidity and flexibility of discrete structures. Rigidity theory, and the related field of kinematics, have a wide range of applications to fields such as material science, robotics, architecture, and computer aided design.

In rigidity theory, rigidity and flexibility are often studied as properties of an underlying combinatorial object. In this thesis, the aim is to study rigidity theoretic problems where the underlying combinatorial object is an incidence geometry. Firstly, we study rigidity problems for realisations of incidence geometries of rank 2 as points and straight lines in the plane. Finding realisations of incidence geometries as points and straight lines in the plane is an interesting problem in its own right that can be formulated as a problem of realisability of rank 3 matroids over the real numbers.

We study motions of rod configurations, which are realisations of incidence geometries as points and straight line segments in the plane, where each line segment is treated as a rigid rod. Specifically, motions of a rod configuration preserve the distance between any two points on a rod. We introduce and investigate a new notion of minimal rigidity for rod configurations. We also prove that rigidity of a rod configuration is equivalent to rigidity of a graph, under certain geometric conditions on the rod configuration. We also find realisations of v₃-configurations that are flexible as rod configurations for ν ≥ 28. We show that all regularrealisations of v₃-configurations for v ≤ 15, and triangle-free v₃-configurations for v ≤ 20 are rigid as rod configurations.

We also consider motions of realisations of incidence geometries as points and straight lines in the plane which preserve only incidences between points and lines. We introduce the notion of projective motions, which are motions of realisations of incidence geometries as points and straight lines in the projective plane which preserve incidences. Furthermore, we introduce the basic tools for investigating rigidity with respectto projective motions. We also investigate the relationship between projective rigidity and higher-order projective rigidity.

Finally, we introduce a sparsity condition on graded posets, and introduce an algorithm which can determine whether a given graded poset satisfies the sparsity condition. We also show that sparsity conditions define a greedoid.

There exists a realism and scalability trade-off in modelling agent-based simulations. As an example, one could create a simpler model with a simpler behavioural model that allows to simulate many agents. However, this simple model can miss important details of the real world. Incorporating these aspects of the real world can increase realism, however, it can come at the cost of scalability. If more aspects of the real world are incorporated into the model, the behavioural model should be tied to these aspects. However, an interdependent behavioural system that considers all information at each time step is usually poorly scalable in terms of deliberative aspects. This poor scalability hinders the expansion of other desirable properties in the model, such as aspects of life and the number of agents.

We propose a context-sensitive deliberation framework that could help increase the scalability of the deliberation without losing behavioural realism. The framework is inspired by Kahneman's concept of thinking fast and thinking slow. On the one hand, it will be capable of fast deliberation that is efficient (scalability). On the other hand, it can sometimes perform slower deliberation that can solve complex situations (behavioural realism). Rather than switching between these two modes, the framework slides, gradually incorporating more information into the deliberation. This is the complexity by need principle. The framework needs to be aware of the context to determine what kind of information to use and what kind of information to deliberate on.

Whether context-sensitive deliberation can increase scalability while retaining realism will be evaluated with a use-case simulation, the Agent-Based Social Simulation of the Coronavirus Crisis (ASSOCC). Context-sensitive deliberation is implemented in the ASSOCC framework. The Original ASSOCC framework is then compared with the context-sensitive ASSOCC variant. The results show that deliberation is no longer the bottleneck, since context-sensitive deliberation achieved a roughly 16-17 times speed-up over the original ASSOCC deliberation model. This speed-up was retained with higher agent numbers, and it can be expected that if deliberation contains more aspects, context-sensitive deliberation will be capable of an even greater speed-up. The behavioural and infection curves were similar between the two models, thus the realism of the model is retained. In conclusion, the work shows that context-sensitive deliberation can increase scalability and retain realism in agent-based simulations.

Background and aims: Symptomatic carotid stenosis cause approxi-mately 15% of all ischemic stroke. Carotid near-occlusion (CNO) is a severe carotid stenosis causing a diameter reduction (collapse) in the distal internal carotid artery (ICA). CNO accounts for approximatelyone-third of cases with symptomatic carotid stenosis. Despite its clinical relevance, CNO remains as an understudied condition in terms of diagnostics, pathophysiology and risk of recurrent stroke.

The overall aim of this thesis is to advance the understanding of underlying pathophysiology behind stroke in CNO, targeting a potentially embolic mechanism. Moreover, we aim to evaluate the short-term risk of strokeand the accuracy of radiologists in identifying CNO using computed tomography angiography (CTA).

Methods: We conducted four studies using retrospective and prospective data from the “Transatlantic Carotid Near-Occlusion Study cohort” (TACNOS) and the “Umeå Carotid Cohort” (UCC). In the TACNOS cohort,we retrospectively reviewed medical records to assess short-term risk ofrecurrent ipsilateral ischemic stroke (study I) and routine CTA reports to evaluate the diagnostic accuracy of radiologists in identifying CNO on routine CTA (study II). In the UCC cohort, we prospectively assessed theincidence of microembolic signals (MES) on transcranial Doppler (TCD) distal to symptomatic CNO (study III), as a biomarker of embolic mechanism. By use of pooled TACNOS and UCC data, we further investigatedthe potentially embolic mechanism in CNO (study IV), by assessing the prevalence of ipsilateral intracranial medium- or large-vessel occlusions, defined as “ipsilateral presumed embolic vessel occlusions” (iPEVO) on CTA among patients with symptomatic CNO.iv

Results: In study I, 99 of 365 included patients had CNO (42 with full collapse and 57 without full collapse). The 90-day risk of recurrent ipsi-lateral ischemic stroke was 30% in symptomatic CNO with full collapse, 22% in symptomatic CNO without full collapse, compared to 15% in symptomatic conventional ≥50% carotid stenosis (p = 0.012). In study II, routine CTA interpretation by radiologists showed strikingly low diagnostic accuracy for carotid near-occlusion, with sensitivity of only 8-22%.

Of 109 cases with TCD-recordings from the UCC cohort, MES incidencein study III was at least similarly high in symptomatic CNO (46%) compared to symptomatic conventional ≥50% carotid stenosis (27%; p = 0.08) and significantly higher than asymptomatic ≥50% carotid stenosis (6%; p = 0.005). Out of 124 cases with CTA on the same day following stroke onset, the prevalence of iPEVO in study IV was higher in symptomatic CNO with full collapse (89%) and without full collapse (44%) than symptomatic conventional ≥50% carotid stenosis (10%; p < 0.001).

Conclusions: CNO is underrecognized in Sweden. The high short-term risk of recurrent stroke in CNO highlights the urgent need for enhanced education and diagnostic radiological methods. An embolic mechanism behind stroke in CNO appears central, the next key step should be to understand the potential role of concurrent hemodynamic factors to guide future treatment strategies.

Cloud computing has transformed the IT landscape by offering users and orga-nizations on-demand access to computing power, storage, data processing, andmachine learning resources. Despite the benefits, cloud resource managementfaces challenges due to the heterogeneous and dynamic nature of workloads.Inefficient provisioning manifests in two critical forms: underprovisioning leadsto degraded Quality of Service (QoS) and unmet Service-Level Agreements(SLAs), while overprovisioning results in unnecessary energy consumption andhigh operational costs. With the current rise of AI and machine learning in-novations, machine learning-based workload prediction for resource provisionplays a vital role in predicting future scenarios and identifying new occurrences,enabling service providers to prepare ahead of time. However, various challengesare associated with machine learning-based workload prediction.This thesis addresses the challenges of machine learning-based workloadprediction in cloud environments, including data drift due to dynamic workloads,high computational overhead, and storage overhead. Firstly, cloud workloads aredynamic, and models trained with old historical data can become obsolete overtime. We addressed the challenge of accurate prediction and data drift by incor-porating machine learning and streaming data processing algorithms to assistadaptive prediction. Secondly, constantly training and updating deep learningmodels adds significant computational overhead to the cloud infrastructure. Weaddressed this problem by proposing a solution that incorporates a knowledgebase repository with transfer learning-based adaptation. Moreover, we exploredthe tradeoff between model accuracy and computational overhead. Finally, wepropose a data compression mechanism that leverages an autoencoder to reducestorage overhead resulting from the continuous generation of monitoring datain cloud management systems.Our findings reveal that the proposed methods have significantly improvedthe machine learning-based cloud management system. Extensive evaluationusing real-world datasets reveals that the proposed methods facilitate thecreation of accurate predictions, even in the face of ever-changing patterns incloud workloads. Moreover, the methods reduced computation overhead byleveraging existing knowledge and highlighting the tradeoff required to achievea balance between prediction accuracy and computation overhead.

The widespread adoption of machine learning (ML) in various domains has enabled the extraction of meaningful insights from complex, large-scale datasets. However, recent research has revealed that ML models are vulnerable to a range of privacy attacks which can expose sensitive information about the individuals in the training data. With regulatory frameworks like the General Data Protection Regulation (GDPR) which enforces strict requirements on data sharing, the need for privacy-preserving solutions has become increasingly critical. As the world becomes more digital, massive volumes of data are generated, often in high-dimensional spaces, where the number of attributes matches or exceeds the number of samples. ML models are extensively used to process such data, making it critical to protect both the data and the models from privacy attacks. 

Traditional anonymization techniques such as k-anonymity and differential privacy often fall short when applied to high-dimensional datasets, because as dimensionality of the data increase, data-points tends to concentrate in the sparse regions of the feature space, making it difficult to find clusters of similar records. Therefore, this thesis proposes a set of privacy-preserving methodologies tailored for high-dimensional data and large-scale foundation models.

In this thesis, we begin by exploring manifold learning techniques to project high-dimensional data into a lower-dimensional latent space while preserving the intrinsic geometric structure of the original data. This transformation enhances the effectiveness of anonymization while maintaining data utility.  Building on this, we then present a novel hybrid privacy method that integrates the strengths of k-anonymity with differential privacy, enabling robust anonymization that preserves both privacy and the underlying data structure. We further investigate synthetic data generation as a privacy-preserving alternative to using sensitive data, leveraging advanced generative models such as GANs and VAEs to produce high-quality synthetic datasets. To enhance the quality of the generated data, we propose techniques that preserve the intrinsic structure of the original high-dimensional data and incorporate prior domain knowledge to guide the generation process. We rigorously evaluate the synthetic data in terms of statistical fidelity, privacy risks, ML utility, and distributional capabilities through detailed visualizations. We then address high-dimensionality and privacy concerns in the context of large-scale foundation models. We propose two model compression strategies using knowledge distillation and pruning, that effectively reduce the number of model parameters while preserving performance and enhancing the privacy of the system. 

Collectively, the thesis contributes towards building privacy-aware AI systems by developing practical solutions that address the complex interplay between high-dimensionality and privacy models. 

In eukaryotes, mitochondria are energy-producing organelles that are unique in containing their own genome, known as mitochondrial DNA (mtDNA). This multicopy DNA encodes 13 essential proteins that are fundamental to energy production. Consequently, defects in mtDNA can disrupt cellular energy processes, potentially leading to disorders ranging from mitochondrial diseases to cancers. Understanding the factors that protect mtDNA and maintain its stability is therefore crucial for ensuring efficient energy production and cellular health.

The incorporation of ribonucleotides (rNMPs) into DNA leads to severe genomic instability, including the formation of DNA breaks. Moreover, rNMP removal through topoisomerase 1 activity can produce harmful intermediates and generate short deletions within repetitive sequences. In this study, we investigate the endoribonuclease activity of mitochondrial topoisomerase 1 (hTop1mt) and compare it to that of its nuclear homolog, hTop1. Using various biochemical assays, we demonstrate that the mitochondrial enzyme exhibits a lower endoribonuclease activity than its nuclear counterpart, a difference that is partially explained by its weaker DNA-binding affinity and slower cleavage kinetics on linear DNA substrates containing an rNMP. Moreover, hTop1mt lacks the preference for cleaving at rNMPs that hTop1 exhibits. Consequently, on repetitive sequences, I showed that hTop1mt generates fewer rNMP-dependent deletions. These findings suggest that the potentially deleterious side effects of this rNMP repair pathway are minimized in mitochondria, providing further insight into why rNMPs in the mitochondrial genome are better tolerated than in the nuclear genome.

Mitochondrial single-stranded DNA-binding protein (mtSSB) is critical for maintaining mtDNA integrity during replication. Mutations in the SSBP1 gene, which encodes mtSSB, are frequently observed—being linked not only to mitochondrial diseases but also to cancer. However, the functional consequences of these cancer-associated mutations are not yet well understood. In this study, we investigated eight cancer-associated SSBP1 mutations through biochemical characterization. I showed that all mutant mtSSB proteins still form tetramers, but their DNA-binding capabilities and thermal stability was compromised. Collectively, these findings suggest that cancer-associated SSBP1 mutations can significantly impair mtSSB function, potentially leading to destabilization of mtDNA integrity.

Horizontal gene transfer between bacteria is closely linked to the spread of antibiotic resistance, and the type IV secretion system (T4SS) plays a crucial role in mediating this process. The pCF10 plasmid from the commensal bacterium Enterococcus faecalis encodes the machinery required for such transfer. In this study, we focused on characterizing the PrgE protein encoded by pCF10, which is proposed to function as a single-stranded DNA-binding (SSB) factor based on its sequence homology with known SSBs. In our structural analysis, we found that PrgE possesses the characteristic OB-fold typical of SSB proteins, yet it displays unusual DNA-binding properties. Specifically, we found that PrgE binds ssDNA very weakly and, surprisingly, exhibits a similar low affinity for dsDNA. In summary, PrgE is an OB-fold protein with atypical DNA interaction properties, and its precise role in the context of horizontal gene transfer remains to be fully elucidated.

In this thesis, I biochemically characterized three distinct DNA-binding proteins from different biological contexts. These findings not only deepen our understanding of the molecular functions of these proteins but could, hopefully, help in the development of novel therapies to enhance human health.