Case-Based Reasoning (CBR) is a symbolic Artificial Intelligence (AI) approach that has been successfully applied across various domains, including medical diagnosis, product configuration, and customer support, to solve problems based on experiential knowledge and analogy. A key aspect of CBR is its problem-solving procedure, where new solutions are created by referencing similar experiences, which makes CBR explainable and effective even with small amounts of data. However, one of the most significant challenges in CBR lies in defining and computing meaningful similarities between new and past problems, which heavily relies on domain-specific knowledge. This knowledge, typically only available through human experts, must be manually acquired, leading to what is commonly known as the knowledge-acquisition bottleneck. One way to mitigate the knowledge-acquisition bottleneck is through a hybrid approach that combines the symbolic reasoning strengths of CBR with the learning capabilities of Deep Learning (DL), a sub-symbolic AI method. DL, which utilizes deep neural networks, has gained immense popularity due to its ability to automatically learn from raw data to solve complex AI problems such as object detection, question answering, and machine translation. While DL minimizes manual knowledge acquisition by automatically training models from data, it comes with its own limitations, such as requiring large datasets, and being difficult to explain, often functioning as a "black box". By bringing together the symbolic nature of CBR and the data-driven learning abilities of DL, a neuro-symbolic, hybrid AI approach can potentially overcome the limitations of both methods, resulting in systems that are both explainable and capable of learning from data. The focus of this thesis is on integrating DL into the core task of similarity assessment within CBR, specifically in the domain of process management. Processes are fundamental to numerous industries and sectors, with process management techniques, particularly Business Process Management (BPM), being widely applied to optimize organizational workflows. Process-Oriented Case-Based Reasoning (POCBR) extends traditional CBR to handle procedural data, enabling applications such as adaptive manufacturing, where past processes are analyzed to find alternative solutions when problems arise. However, applying CBR to process management introduces additional complexity, as procedural cases are typically represented as semantically annotated graphs, increasing the knowledge-acquisition effort for both case modeling and similarity assessment. The key contributions of this thesis are as follows: It presents a method for preparing procedural cases, represented as semantic graphs, to be used as input for neural networks. Handling such complex, structured data represents a significant challenge, particularly given the scarcity of available process data in most organizations. To overcome the issue of data scarcity, the thesis proposes data augmentation techniques to artificially expand the process datasets, enabling more effective training of DL models. Moreover, it explores several deep learning architectures and training setups for learning similarity measures between procedural cases in POCBR applications. This includes the use of experience-based Hyperparameter Optimization (HPO) methods to fine-tune the deep learning models. Additionally, the thesis addresses the computational challenges posed by graph-based similarity assessments in CBR. The traditional method of determining similarity through subgraph isomorphism checks, which compare nodes and edges across graphs, is computationally expensive. To alleviate this issue, the hybrid approach seeks to use DL models to approximate these similarity calculations more efficiently, thus reducing the computational complexity involved in graph matching. The experimental evaluations of the corresponding contributions provide consistent results that indicate the benefits of using DL-based similarity measures and case retrieval methods in POCBR applications. The comparison with existing methods, e.g., based on subgraph isomorphism, shows several advantages but also some disadvantages of the compared methods. In summary, the methods and contributions outlined in this work enable more efficient and robust applications of hybrid CBR and DL in process management applications.

The goal of this work is to compare operators that are defined on probably varying Hilbert spaces. Distance concepts for operators as well as convergence concepts for such operators are explained and examined. For distance concepts we present three main notions. All have in common that they use space-linking operators that connect the spaces. At first, we look at unitary maps and compare the unitary orbits of the operators. Then, we consider isometric embeddings, which is based on a concept of Joachim Weidmann. Then we look at contractions but with more norm equations in comparison. The latter idea is based on a concept of Olaf Post called quasi-unitary equivalence. Our main result is that the unitary and isometric distances are equal provided the operators are both self-adjoint and have 0 in their essential spectra. In the third chapter, we focus specifically on the investigation of these distance terms for compact operators or operators in p-Schatten classes. In this case, the interpretation of the spectra as null sequences allows further distance investigation. Chapter four deals mainly with convergence terms of operators on varying Hilbert spaces. The analyses in this work deal exclusively with concepts of norm resolvent convergence. The main conclusion of the chapter is that the generalisation for norm resolvent convergence of Joachim Weidmann and the generalisation of Olaf Post, called quasi-unitary equivalence, are equivalent to each other. In addition, we specify error bounds and deal with the convergence speed of both concepts. Two important implications of these convergence notions are that the approximation is spectrally exact, i.e., the spectra converge suitably, and that the convergence is transferred to the functional calculus of the bounded functions vanishing at infinity.

Globalization significantly transforms labor markets. Advances in production technologies, transportation, and political integration reshape how and where goods and services are produced. Local economic conditions and diverse policy responses create varying speeds of change, affecting regions' attractiveness for living and working -- and promoting mobility. Competition for talent necessitates a deep understanding of why individuals choose specific destinations, how to ensure their effective labor market integration, and what workplace factors affect workers' well-being. This thesis focuses on two crucial aspects of labor market change -- Migration and workplace technological change. It contributes to our understanding of the determinants of labor mobility, the factors facilitating migrant integration, and the role of workplace automation for worker well-being. Chapter 2 investigates the relationship between minimum wages (MWs) and regional worker mobility in the EU. EU citizens are free to work anywhere in the common market, which allows them to take advantage of the significant variation in MWs across the EU. However, although MWs are set at the national level, it is also their local relevance that varies substantially -- depending on factors such as the share of affected workers or the extent to which they shift local compensation levels. These variations may attract workers from elsewhere, from within a country or from abroad. Analyzing regional variations in the Kaitz index, a measure of local MW impact, reveals that higher MWs can significantly increase inflows of low-skilled EU workers, particularly in central Europe. Chapter 3 examines the inequality in returns to skills experienced by immigrants, focusing on the role of linguistic proximity between migrants' origin and destination countries. Harmonized individual-level data from nine linguistically diverse migrant-hosting economies allows for an analysis of the wage gaps faced by immigrants from various origins, implicitly indicating how well they and their skills are integrated into the local labor markets. The analysis reveals that greater linguistic distance is associated with a higher wage penalty for highly skilled immigrants and a lower position in the wage distribution for those without tertiary education. Chapter 4 investigates an institutional factor potentially relevant for the integration of immigrants -- the labor market impact of Confucius Institutes (CIs), Chinese government-sponsored institutions that promote Chinese language and culture abroad. CIs have been found to foster trade and cultural exchange, indicating their potential relevance in shaping attitudes and trust of natives towards China and Chinese individuals. Examining the relationship between local CI presence and the wages of Chinese immigrants in local labor markets of the United States, the analysis reveals that CIs associate with significantly reduced wages for nearby residing Chinese immigrants. An event study demonstrates that the mere announcement of a new CI negatively impacts local wages for Chinese immigrants, independent of the CI's actual opening. Chapter 5 explores how working in automatable jobs affects life satisfaction in Germany. Following earlier literature, we classify occupations by potential for automation, and define the top third of occupations in this metric as \textit{automatable jobs}. We find workers in highly automatable jobs reporting a lower life satisfaction. Moreover, we detect a non-linearity, where workers in moderately automatable jobs (the second third of the distribution) experience a positive association with life satisfaction. Overall, the negative relationship of automation is most pronounced among younger and blue-collar workers, irrespective of the non-linearity.

This dissertation addresses the measurement and evaluation of the energy and resource efficiency of software systems. Studies show that the environmental impact of Information and Communications Technologies (ICT) is steadily increasing and is already estimated to be responsible for 3 % of the total greenhouse gas (GHG) emissions. Although it is the hardware that consumes natural resources and energy through its production, use, and disposal, software controls the hardware and therefore has a considerable influence on the used capacities. Accordingly, it should also be attributed a share of the environmental impact. To address this softwareinduced impact, the focus is on the continued development of a measurement and assessment model for energy and resource-efficient software. Furthermore, measurement and assessment methods from international research and practitioner communities were compared in order to develop a generic reference model for software resource and energy measurements. The next step was to derive a methodology and to define and operationalize criteria for evaluating and improving the environmental impact of software products. In addition, a key objective is to transfer the developed methodology and models to software systems that cause high consumption or offer optimization potential through economies of scale. These include, e. g., Cyber-Physical Systems (CPS) and mobile apps, as well as applications with high demands on computing power or data volumes, such as distributed systems and especially Artificial Intelligence (AI) systems. In particular, factors influencing the consumption of software along its life cycle are considered. These factors include the location (cloud, edge, embedded) where the computing and storage services are provided, the role of the stakeholders, application scenarios, the configuration of the systems, the used data, its representation and transmission, or the design of the software architecture. Based on existing literature and previous experiments, distinct use cases were selected that address these factors. Comparative use cases include the implementation of a scenario in different programming languages, using varying algorithms, libraries, data structures, protocols, model topologies, hardware and software setups, etc. From the selection, experimental scenarios were devised for the use cases to compare the methods to be analyzed. During their execution, the energy and resource consumption was measured, and the results were assessed. Subtracting baseline measurements of the hardware setup without the software running from the scenario measurements makes the software-induced consumption measurable and thus transparent. Comparing the scenario measurements with each other allows the identification of the more energyefficient setup for the use case and, in turn, the improvement/optimization of the system as a whole. The calculated metrics were then also structured as indicators in a criteria catalog. These indicators represent empirically determinable variables that provide information about a matter that cannot be measured directly, such as the environmental impact of the software. Together with verification criteria that must be complied with and confirmed by the producers of the software, this creates a model with which the comparability of software systems can be established. The gained knowledge from the experiments and assessments can then be used to forecast and optimize the energy and resource efficiency of software products. This enables developers, but also students, scientists and all other stakeholders involved in the life cycleof software, to continuously monitor and optimize the impact of their software on energy and resource consumption. The developed models, methods, and criteria were evaluated and validated by the scientific community at conferences and workshops. The central outcomes of this thesis, including a measurement reference model and the criteria catalog, were disseminated in academic journals. Furthermore, the transfer to society has been driven forward, e. g., through the publication of two book chapters, the development and presentation of exemplary best practices at developer conferences, collaboration with industry, and the establishment of the eco-label “Blue Angel” for resource and energy-efficient software products. In the long term, the objective is to effect a change in societal attitudes and ultimately to achieve significant resource savings through economies of scale by applying the methods in the development of software in general and AI systems in particular.

Although universality has fascinated over the last decades, there are still numerous open questions in this field that require further investigation. In this work, we will mainly focus on classes of functions whose Fourier series are universal in the sense that they allow us to approximate uniformly any continuous function defined on a suitable subset of the unit circle. The structure of this thesis is as follows. In the first chapter, we will initially introduce the most important notation which is needed for our following discussion. Subsequently, after recalling the notion of universality in a general context, we will revisit significant results concerning universality of Taylor series. The focus here is particularly on universality with respect to uniform convergence and convergence in measure. By a result of Menshov, we will transition to universality of Fourier series which is the central object of study in this work. In the second chapter, we recall spaces of holomorphic functions which are characterized by the growth of their coefficients. In this context, we will derive a relationship to functions on the unit circle via an application of the Fourier transform. In the second part of the chapter, our attention is devoted to the $\mathcal{D}_{\textup{harm}}^p$ spaces which can be viewed as the set of harmonic functions contained in the $W^{1,p}(\D)$ Sobolev spaces. In this context, we will also recall the Bergman projection. Thanks to the intensive study of the latter in relation to Sobolev spaces, we can derive a decomposition of $\mathcal{D}_{\textup{harm}}^p$ spaces which may be seen as analogous to the Riesz projection for $L^p$ spaces. Owing to this result, we are able to provide a link between $\mathcal{D}_{\textup{harm}}^p$ spaces and spaces of holomorphic functions on $\mathbb{C}_\infty \setminus \s$ which turns out to be a crucial step in determining the dual of $\mathcal{D}_{\textup{harm}}^p$ spaces. The last section of this chapter deals with the Cauchy dual which has a close connection to the Fantappié transform. As an application, we will determine the Cauchy dual of the spaces $D_\alpha$ and $D_{\textup{harm}}^p$, two results that will prove to be very helpful later on. Finally, we will provide a useful criterion that establishes a connection between the density of a set in the direct sum $X \oplus Y$ and the Cauchy dual of the intersection of the respective spaces. The subsequent chapter will delve into the theory of capacities and, consequently, potential theory which will prove to be essential in formulating our universality results. In addition to introducing further necessary terminologies, we will define capacities in the first section following [16], however in the frame of separable metric spaces, and revisit the most important results about them. Simultaneously, we make preparations that allow us to define the $\mathrm{Li}_\alpha$-capacity which will turn out to be equivalent to the classical Riesz $\alpha$-capacity. The $\mathrm{Li}_\alpha$-capacity proves to be more adapted to the $D_\alpha$ spaces. It becomes apparent in the course of our discussion that the $\mathrm{Li}_\alpha$-capacity is essential to prove uniqueness results for the class $D_\alpha$. This leads to the centerpiece of this chapter which forms the energy formula for the $\mathrm{Li}_\alpha$-capacity on the unit circle. More precisely, this identity establishes a connection between the energy of a measure and its corresponding Fourier coefficients. We will briefly deal with the complement-equivalence of capacities before we revisit the concept of Bessel and Riesz capacities, this time, however, in a much more general context, where we will mainly rely on [1]. Since we defined capacities on separable metric spaces in the first section, we can draw a connection between Bessel capacities and $\mathrm{Li}_\alpha$-capacities. To conclude this chapter, we would like to take a closer look at the geometric meaning of capacities. Here, we will point out a connection between the Hausdorff dimension and the polarity of a set, and transfer it to the $\mathrm{Li}_\alpha$-capacity. Another aspect will be the comparison of Bessel capacities across different dimensions, in which the theory of Wolff potentials crystallizes as a crucial auxiliary tool. In the fourth chapter of this thesis, we will turn our focus to the theory of sets of uniqueness, a subject within the broader field of harmonic analysis. This theory has a close relationship with sets of universality, a connection that will be further elucidated in the upcoming chapter. The initial section of this chapter will be dedicated to the notion of sets of uniqueness that is specifically adapted to our current context. Building on this concept, we will recall some of the fundamental results of this theory. In the subsequent section, we will primarily rely on techniques from previous chapters to determine the closed sets of uniqueness for the class $\mathcal{D}_{\alpha}$. The proofs we will discuss are largely influenced by [16, p.\ 178] and [9, pp.\ 82]. One more time, it will become evident that the introduction of the $\mathrm{Li}_\alpha$-capacity in the third chapter and the closely associated energy formula on the unit circle, were the pivotal factors that enabled us to carry out these proofs. In the final chapter of our discourse, we will present our results on universality. To begin, we will recall a version of the universality criterion which traces back to the work of Grosse-Erdmann (see [26]). Coupled with an outcome from the second chapter, we will prove a result that allows us to obtain the universality of a class using the technique of simultaneous approximation. This tool will play a key role in the proof of our universality results which will follow hereafter. Our attention will first be directed toward the class $D_\alpha$ with $\alpha$ in the interval $(0,1]$. Here, we summarize that universality with respect to uniform convergence occurs on closed and $\alpha$-polar sets $E \subset \s$. Thanks to results of Carleson and further considerations, which particularly rely on the favorable behavior of the $\mathrm{Li}_\alpha$-kernel, we also find that this result is sharp. In particular, it may be seen as a generalization of the universality result for the harmonic Dirichlet space. Following this, we will investigate the same class, however, this time for $\alpha \in [-1,0)$. In this case, it turns out that universality with respect to uniform convergence occurs on closed and $(-\alpha)$-complement-polar sets $E \subset \s$. In particular, these sets of universality can have positive arc measure. In the final section, we will focus on the class $D_{\textup{harm}}^p$. Here, we manage to prove that universality occurs on closed and $(1,p)$-polar sets $E \subset \s$. Through results of Twomey [68] combined with an observation by Girela and Pélaez [23], as well as the decomposition of $D_{\textup{harm}}^p$, we can deduce that the closed sets of universality with respect to uniform convergence of the class $D_{\textup{harm}}^p$ are characterized by $(1,p)$-polarity. We conclude our work with an application of the latter result to the class $D^p$. We will show that the closed sets of divergence for the class $D^p$ are given by the $(1,p)$-polar sets.

This thesis deals with consumption-investment allocation problems with Epstein-Zin recursive utility, building upon the dualization procedure introduced by [Matoussi and Xing, 2018]. While their work exclusively focuses on truly recursive utility, we extend their procedure to include time-additive utility using results from general convex analysis. The dual problem is expressed in terms of a backward stochastic differential equation (BSDE), for which existence and uniqueness results are established. In this regard, we close a gap left open in previous works, by extending results restricted to specific subsets of parameters to cover all parameter constellations within our duality setting. Using duality theory, we analyze the utility loss of an investor with recursive preferences, that is, her difference in utility between acting suboptimally in a given market, compared to her best possible (optimal) consumption-investment behaviour. In particular, we derive universal power utility bounds, presenting a novel and tractable approximation of the investors’ optimal utility and her welfare loss associated to specific investment-consumption choices. To address quantitative shortcomings of those power utility bounds, we additionally introduce one-sided variational bounds that offer a more effective approximation for recursive utilities. The theoretical value of our power utility bounds is demonstrated through their application in a new existence and uniqueness result for the BSDE characterizing the dual problem. Moreover, we propose two approximation approaches for consumption-investment optimization problems with Epstein-Zin recursive preferences. The first approach directly formalizes the classical concept of least favorable completion, providing an analytic approximation fully characterized by a system of ordinary differential equations. In the special case of power utility, this approach can be interpreted as a variation of the well-known Campbell-Shiller approximation, improving some of its qualitative shortcomings with respect to state dependence of the resulting approximate strategies. The second approach introduces a PDE-iteration scheme, by reinterpreting artificial completion as a dynamic game, where the investor and a dual opponent interact until reaching an equilibrium that corresponds to an approximate solution of the investors optimization problem. Despite the need for additional approximations within each iteration, this scheme is shown to be quantitatively and qualitatively accurate. Moreover, it is capable of approximating high dimensional optimization problems, essentially avoiding the curse of dimensionality and providing analytical results.

In bilevel optimization, some of the variables of an optimization problem have to be an optimal solution to another nested optimization problem. This specific structure renders bilevel optimization a powerful tool for modeling hierarchical decision-making processes, which arise in various real-world applications such as in critical infrastructure defense, transportation, or energy. Due to their nested structure, however, bilevel problems are also inherently hard to solve—both in theory and in practice. Further challenges arise if, e.g., bilevel problems under uncertainty are considered. In this dissertation, we address different types of uncertainties in bilevel optimization using techniques from robust optimization. We study mixed-integer linear bilevel problems with lower-level objective uncertainty, which we tackle using the notion of Gamma-robustness. We present two exact branch-and-cut approaches to solve these Gamma-robust bilevel problems, along with cuts tailored to the important class of monotone interdiction problems. Given the overall hardness of the considered problems, we additionally propose heuristic approaches for mixed-integer, linear, and Gamma-robust bilevel problems. The latter rely on solving a linear number of deterministic bilevel problems so that no problem-specific tailoring is required. We assess the performance of both the exact and the heuristic approaches through extensive computational studies. In addition, we study the problem of determining optimal tolls in a traffic network in which the network users hedge against uncertain travel costs in a robust way. The overall toll-setting problem can be seen as a single-leader multi-follower problem with multiple robustified followers. We model this setting as a mathematical problem with equilibrium constraints, for which we present a mixed-integer, nonlinear, and nonconvex reformulation that can be tackled using state-of-the-art general-purpose solvers. We further illustrate the impact of considering robustified followers on the toll-setting policies through a case study. Finally, we highlight that the sources of uncertainty in bilevel optimization are much richer compared to single-level optimization. To this end, we study two aspects related to so-called decision uncertainty. First, we propose a strictly robust approach in which the follower hedges against erroneous observations of the leader's decision. Second, we consider an exemplary bilevel problem with a continuous but nonconvex lower level in which algorithmic necessities prevent the follower from making a globally optimal decision in an exact sense. The example illustrates that even very small deviations in the follower's decision may lead to arbitrarily large discrepancies between exact and computationally obtained bilevel solutions.

Based on data collected from two surveys conducted in Germany and Taiwan, my first paper (Chapter 2) examines the impact of culture through language priming (Chinese vs. German or English) on individuals’ price fairness perception and attitudes towards government intervention and economic policy involving inequality. We document large cross-language differences: in both surveys, subjects who were asked and answered in Chinese demonstrated significantly higher perceived price fairness in a free market mechanism than their counterparts who completed the survey in German or English language. They were also more inclined to accept a Pareto improvement policy which increases social and economic inequality. In the second survey, Chinese language induced also a lower readiness to accept government intervention in markets with price limits compared to English language. Since language functions as a cultural mindset prime, our findings imply that culture plays an important role in fairness perception and preferences regarding social and economic inequality. Chapter 3 of this work deals with patriotism priming. By conducting two online experimental studies conducted in Germany and China, we tested three different kinds of priming methods for constructive and blind patriotism respectively. Subjects were randomly distributed to one of three treatments motivated by previous studies in different countries: a constructive patriotism priming treatment, a blind patriotism priming treatment and a non-priming baseline. While the first experiment had a between-subject design, the second one enabled both a between-subject and within-subject comparison, since the level of patriotism of individuals was measured before and after priming respectively. The design of the second survey also enabled a comparison among the three priming methods for constructive and blind patriotism. The results showed that the tested methods, especially the national achievements as a priming mechanism, functioned well overall for constructive patriotism. Surprisingly, the priming for blind patriotism did not work in either Germany or China and the opposite results were observed. Discussion and implications for future studies are provided at the end of the chapter. Using data from the same studies as in Chapter 3, Chapter 4 examines the impact of patriotism on individuals’ fairness perception and preferences regarding inequality and on their attitudes toward economic policy involving inequality. Across surveys and countries, a positive and significant effect of blind patriotism on economic individualism was found. For China, we also found a significant relationship between blind patriotism and the agreement to unequal economic policy. In contrast to blind patriotism, we did not find an association of constructive patriotism to economic individualism and to attitudes toward economic policy involving inequality. Political and economic implications based on the results are discussed. The last chapter (Chapter 5) studies the self-serving bias (when an individual’s perception about fairness is biased by self-interest) in the context of price setting and profit distribution. By analyzing data from four surveys conducted in six countries, we found that the stated appropriate product price and the fair allocation of profit was significantly higher, when the outcome was favorable to oneself. This self-serving bias in price fairness perception, however, differed across countries significantly and was significantly higher in Germany, Taiwan and China than in Vietnam, Estonia and Japan. Although economic individualism and masculinity were found to have a significant negative effect on self-interest bias in price fairness judgment, they did not sufficiently explain the differences in self-interest bias between countries. Furthermore, we also observed an increase of self-interest bias in profit allocation over time in time-series data for one country (Germany) with data from 2011 to 2023. The four papers are all co-authored with Prof. Marc Oliver Rieger, and the first paper has been accepted for publications in Review of Behavioral Economics.

This thesis contains three parts that are all connected by their contribution to research about the effects of trading apps on investment behavior. The primary motivation for this study is to investigate the previously undetermined consequences and effects of trading apps, which are a new phenomenon in the broker market, on the investment and risk behavior of Neobroker users. Chapter 2 addresses the characteristics of a typical Neobroker user and a former Neobroker user and the impact of trading apps on the investment and risk behavior of their users. The results show that Neobroker users are significantly more risk tolerant than the general German population and are influenced by trading apps regarding their investment and risk behavior. Low trading fees and the low minimum investment amount are the main reasons for the use of trading apps. Investors who stop using trading apps mostly stop investing altogether. Another worrying result is that financial literacy among all groups is low and most Neobroker users have wrong conceptions about how trading apps earn money. In general, the financial literacy of all groups considered in this chapter is surprisingly low. The third chapter investigates the effects of trading apps on investment behavior over time and compares the investment and risk behavior of Neobroker users and general investors. By using representative data of German Neobroker users, who were surveyed repeatedly over a 8-month time interval, it becomes possible to determine causal effects of the use of trading apps over time. In total, the financial literacy of Neobroker users increases with the longer use of a trading app. A worrying result is that the risk tolerance of Neobroker users rises significantly over time. Male Neobroker users gain a higher annual return (non-risk-adjusted) than female Neobroker users. In comparison to general investors, Neobroker users are significantly younger, more risk tolerant, more likely to buy derivatives and gain a higher annual return (non-risk-adjusted). The fourth chapter analyses the impact of personality traits on the investment and risk behavior of Neobroker users. The results show that the BIG-5 personality traits have an impact on the investment behavior of Neobroker users. Two personality traits, openness and conscientiousness, stand out the most, as these two have explanatory power over various aspects of the behavior of Neobroker users. In particular, whether they buy different financial products than planned, the time they inform themselves about financial markets, the variety of financial products owned, and the reasons to use a Neobroker. Surprisingly, the risk tolerance of Neobroker users and the reasons to invest are not connected to any personal dimension. Whether a participant uses a trading app or a traditional broker to invest is respectively influenced by different personality traits.

In machine learning, classification is the task of predicting a label for each point within a data set. When the class of each point in the labeled subset is already known, this information is used to recognize patterns and make predictions about the points in the remainder of the set, referred to as the unlabeled set. This scenario falls in the field of supervised learning. However, the number of labeled points can be restricted, because, e.g., it is expensive to obtain this information. Besides, this subset may be biased, such as in the case of self-selection in a survey. Consequently, the classification performance for unlabeled points may be limited. To improve the reliability of the results, semi-supervised learning tackles the setting of labeled and unlabeled data. Moreover, in many cases, additional information about the size of each class can be available from undisclosed sources. This cumulative thesis presents different studies to combine this external cardinality constraint information within three important algorithms for binary classification in the supervised context: support vector machines (SVM), classification trees, and random forests. From a mathematical point of view, we focus on mixed-integer programming (MIP) models for semi-supervised approaches that consider a cardinality constraint for each class for each algorithm. Furthermore, since the proposed MIP models are computationally challenging, we also present techniques that simplify the process of solving these problems. In the SVM setting, we introduce a re-clustering method and further computational techniques to reduce the computational cost. In the context of classification trees, we provide correct values for certain bounds that play a crucial role for the solver performance. For the random forest model, we develop preprocessing techniques and an intuitive branching rule to reduce the solution time. For all three methods, our numerical results show that our approaches have better statistical performances for biased samples than the standard approach.