Fakultät für Mathematik, Informatik und Statistik - Digitale Hochschulschriften der LMU - Teil 02/02

Credit scoring models are the basis for financial institutions like retail and consumer credit banks. The purpose of the models is to evaluate the likelihood of credit applicants defaulting in order to decide whether to grant them credit. The area under the receiver operating characteristic (ROC) curve (AUC) is one of the most commonly used measures to evaluate predictive performance in credit scoring. The aim of this thesis is to benchmark different methods for building scoring models in order to maximize the AUC. While this measure is used to evaluate the predictive accuracy of the presented algorithms, the AUC is especially introduced as direct optimization criterion.

The logistic regression model is the most widely used method for creating credit scorecards and classifying applicants into risk classes. Since this development process, based on the logit model, is standard in the retail banking practice, the predictive accuracy of this proceeding is used for benchmark reasons throughout this thesis.

The AUC approach is a main task introduced within this work. Instead of using the

maximum likelihood estimation, the AUC is considered as objective function to optimize it directly. The coefficients are estimated by calculating the AUC measure with Wilcoxon-Mann-Whitney and by using the Nelder-Mead algorithm for the optimization. The AUC optimization denotes a distribution-free approach, which is analyzed within a simulation study for investigating the theoretical considerations. It can be shown that the approach still works even if the underlying distribution is not logistic.

In addition to the AUC approach and classical well-known methods like generalized additive models, new methods from statistics and machine learning are evaluated for the credit scoring case. Conditional inference trees, model-based recursive partitioning methods and random forests are presented as recursive partitioning algorithms. Boosting algorithms are also explored by additionally using the AUC as a loss function.

The empirical evaluation is based on data from a German bank. From the application scoring, 26 attributes are included in the analysis. Besides the AUC, different performance measures are used for evaluating the predictive performance of scoring models. While classification trees cannot improve predictive accuracy for the current credit scoring case, the AUC approach and special boosting methods provide outperforming results compared to the robust classical scoring models regarding the predictive performance with the AUC measure.

Treatment efficacy in clinical trials is often assessed by time from treatment initiation to occurrence of a certain critical or beneficial event. In most cases the event of interest cannot be observed for all patients, as patients are only followed for a limited time or contact to patients is lost during their follow-up time. Therefore, certain methods were developed in the framework of the so called time-to-event or survival analysis, in order to obtain valid and consistent estimates in the presence of these "censored observations", using all available information. In classical event time analysis only one endpoint exists, as the death of a patient. As patients can die from different causes, in some clinical trials time to one out of two or more mutually exclusive types of event may be of interest. In many oncological studies, for example, time to cancer-specific death is considered as primary endpoint with deaths from other causes acting as so called competing risks. Different methods for data analysis in the competing risks framework were developed in recent years, which either focus on modelling the cause-specific or the subdistribution hazard rate or split the joint distribution of event times and event types into quantities, that can be estimated from observable data. In this work the analysis of event time data in the presence of competing risks is described, including the presentation and discussion of different regression approaches. A major topic of this work is the estimation of cause-specific and subdistribution hazard rates from a mixture model and a new approach using penalized B-splines (P-splines) for estimation of conditional hazard rates in a mixture model is proposed. In order to evaluate the behaviour of the new approach, a simulation study was conducted, using simulation techniques for competing risks data, which are described in detail in this work. The presented regression models were applied to data from a clinical cohort study investigating a risk stratification for cardiac mortality in patients, that survived a myocardial infarction. Finally, the use of the presented methods for event time analysis in the presence of competing risks and results obtained from the simulation study and the data analysis are discussed.

In this work, we derive the time-dependent Hartree(-Fock) equations as an effective dynamics for fermionic many-particle systems. Our main results are the first for a quantum mechanical mean-field dynamics for fermions; in previous works, the mean-field limit is usually either coupled to a semiclassical limit, or the interaction is scaled down so much, that the system behaves freely for large particle number N. We mainly consider systems with total kinetic energy bounded by const N and long-range interaction potentials, e.g., Coulomb interaction. Examples for such systems are large molecules or certain solid states. Our analysis also applies to attractive interactions, as, e.g., in fermionic stars. The fermionic Hartree(-Fock) equations are a standard tool to describe, e.g., excited states or chemical reactions of large molecules (like proteins). A deeper understanding of these equations as an approximation to the time evolution of a many body quantum system is thus highly relevant. We consider the fermionic Hatree equations (i.e., the Hartree-Fock equations without exchange term) in this work, since the exchange term is subleading in our setting. The main result is that the fermionic Hartree dynamics approximates the Schrödinger dynamics well for large N. This statement becomes exact in the thermodynamic limit N to infinity. We give explicit values for the rates of convergence. We prove two types of results. The first type is very general and concerns arbitrary free Hamiltonians (e.g., relativistic, non-relativistic, with external fields) and arbitrary interactions. The theorems give explicit conditions on the solutions to the fermonic Hartree equations under which a derivation of the mean-field dynamics succeeds. The second type of results scrutinizes situations where the conditions are fulfilled. These results are about non-relativistic free Hamiltonians with external fields, systems with total kinetic energy bounded by const N and with long-range interactions of the form x^(-s), with 0 < s < 6/5 (sometimes, for technical reasons, with a weaker or cut off singularity).

We prove our main results by using a new method for deriving mean-field dynamics developed by Pickl in [Lett. Math. Phys., 97(2):151-164, 2011]. This method has been applied successfully in quantum mechanics for deriving the bosonic Hartree and Gross-Pitaevskii equations. Its application to fermions in this work is new. The method is based on a functional that "counts the number of particles outside the condensate", i.e., in the case of fermions, it measures those parts of the Schrödinger wave function that are not in the antisymmetric product of the Hartree states. We show that convergence of the functional to zero (which means that the mean-field equations approximate the dynamics well) is equivalent to convergence of the corresponding reduced one-particle density matrices in trace norm and in Hilbert-Schmidt norm. Finally, we show how also the recently treated semiclassical mean-field limits can be derived with this method.

Semantic search engines aim at improving conventional search with semantic information, or meta-data, on the data searched for and/or on the searchers. So far, approaches to semantic search exploit characteristics of the searchers like age, education, or spoken language for selecting and/or ranking search results. Such data allow to build up a semantic search engine as an extension

of a conventional search engine. The crawlers of well established search engines like Google, Yahoo! or Bing can index documents but, so far, their capabilities to recognize the intentions of searchers are still rather limited. Indeed, taking into account characteristics of the searchers considerably extend both, the quantity of data to analyse and the dimensionality of the search problem. Well established search engines therefore still focus on general search, that is, "search for all", not on specialized search, that is, "search for a few".

This thesis reports on techniques that have been adapted or conceived, deployed, and tested for building a semantic search engine for the very specific context of artworks. In contrast to, for example, the interpretation of X-ray images, the interpretation of artworks is far from being fully automatable. Therefore artwork interpretation has been based on Human Computation, that is, a software-based gathering of contributions by many humans.

The approach reported about in this thesis first relies on so called Games With A Purpose, or GWAPs, for this gathering: Casual games provide an incentive for a potentially unlimited community of humans to contribute with their appreciations of artworks. Designing convenient incentives is less trivial than it might seem at first. An ecosystem of games is needed so as to collect the meta-data on artworks intended for. One game generates the data that can serve as input of another game. This results in semantically rich meta-data that can be used for building up a successful semantic search engine. Thus, a first part of this thesis reports on a "game ecosystem" specifically designed from one known game and including several novel games belonging to the following game classes: (1) Description Games for collecting obvious and trivial meta-data, basically the well-known ESP (for extra-sensorial perception) game of Luis von Ahn, (2) the Dissemination Game Eligo generating translations, (3) the Diversification Game Karido aiming at sharpening differences between the objects, that is, the artworks, interpreted and (3) the Integration Games Combino, Sentiment and TagATag that generate structured meta-data.

Secondly, the approach to building a semantic search engine reported about in this thesis relies on Higher-Order Singular Value Decomposition (SVD). More precisely, the data and meta-data on artworks gathered with the afore mentioned GWAPs are collected in a tensor, that is a mathematical structure generalising matrices to more than only two dimensions, columns and rows. The dimensions considered are the artwork descriptions, the players, and the artwork themselves. A Higher-Order SVD of this tensor is first used for noise reduction in

This thesis reports also on deploying a Higher-Order LSA. The parallel Higher-Order SVD algorithm applied for the Higher-Order LSA and its implementation has been validated on an application related to, but independent from, the semantic search engine for artworks striven for: image compression. This thesis reports on the surprisingly good image compression which can be achieved with Higher-Order SVD. While compression methods based on matrix SVD for each color, the approach reported about in this thesis relies on one single (higher-order) SVD of the whole tensor. This results in both, better quality of the compressed image and in a

significant reduction of the memory space needed.

Higher-Order SVD is extremely time-consuming what calls for parallel computation. Thus, a step towards automatizing the construction of a semantic search

Computer security is a very technical topic that is in many cases hard to grasp for the average user. Especially when using the Internet, the biggest network connecting computers globally together, security and safety are important. In many cases they can be achieved without the user's active participation: securely storing user and customer data on Internet servers is the task of the respective company or service provider, but there are also a lot of cases where the user is involved in the security process, especially when he or she is intentionally attacked. Socially engineered phishing attacks are such a security issue were users are directly attacked to reveal private data and credentials to an unauthorized attacker. These types of attacks are the main focus of the research presented within my thesis.

I have a look at how these attacks can be counteracted by detecting them in the first place but also by mediating these detection results to the user. In prior research and development these two areas have most often been regarded separately, and new security measures were developed without taking the final step of interacting with the user into account. This interaction mainly means presenting the detection results and receiving final decisions from the user. As an overarching goal within this thesis I look at these two aspects united, stating the overall protection as the sum of detection and "user intervention".

Within nine different research projects about phishing protection this thesis gives answers to ten different research questions in the areas of creating new phishing detectors (phishing detection) and providing usable user feedback for such systems (user intervention): The ten research questions cover five different topics in both areas from the definition of the respective topic over ways how to measure and enhance the areas to finally reasoning about what is making sense. The research questions have been chosen to cover the range of both areas and the interplay between them. They are mostly answered by developing and evaluating different prototypes built within the projects that cover a range of human-centered detection properties and evaluate how well these are suited for phishing detection. I also take a look at different possibilities for user intervention (e.g. how should a warning look like? should it be blocking or non-blocking or perhaps even something else?). As a major contribution I finally present a model that combines phishing detection and user intervention and propose development and evaluation recommendations for similar systems. The research results show that when developing security detectors that yield results being relevant for end users such a detector can only be successful in case the final user feedback already has been taken into account during the development process.

Regression models with lagged covariate effects are often used in biostatistical and geo-

physical data analysis.

In the difficult and all-important subject of earthquake research, strong long-lasting

rainfall is assumed to be one of many complex trigger factors that lead to earthquakes.

Geophysicists interpret the rain effect with an increase of pore pressure due to the infiltra-

tion of rain water over a long time period. Therefore, a sensible statistical regression model

examining the influence of rain on the number of earthquakes on day t has to contain rain

information of day t and of preceding days t − 1 to t − L.

In the first part of this thesis, the specific shape of lagged rain influence on the number of

earthquakes is modeled. A novel penalty structure for interpretable and flexible estimates

of lag coefficients based on spline representations is presented. The penalty structure

enables smoothness of the resulting lag course and a shrinkage towards zero of the last

lag coefficient via a ridge penalty. This additional ridge penalty offers an approach to

another problem neglected in previous work. With the help of the additional ridge penalty,

a suboptimal choice of the lag length L is no longer critical. We propose the use of longer

lags, as our simulations indicate that superfluous coefficients are correctly estimated close

to zero.

We provide a user-friendly implementation of our flexible distributed lag (FDL) ap-

proach, that can be used directly in the established R package mgcv for estimation of

generalized additive models. This allows our approach to be immediately included in com-

plex additive models for generalized responses even in hierarchical or longitudinal data

settings, making use of established stable and well-tested algorithms. We demonstrate

the performance and utility of the proposed flexible distributed lag model in a case study

on (micro-) earthquake data from Mount Hochstaufen, Bavaria with focus on the specific

shape of the lagged rain influence on the occurrence of earthquakes in different depths.

The complex meteorological and geophysical data set was collected and provided by the

Geophysical Observatory of the Ludwig-Maximilians University Munich. The benefit of

flexible distributed lag modeling is shown in a detailed simulation study.

In the second part of the thesis, the penalization concept is extended to lagged non-

linear covariate influence. Here, we extend an approach of Gasparrini et al. (2010), that

was up to now unpenalized. Detailed simulation studies illustrate again the benefits of

the penalty structure. The flexible distributed lag nonlinear model is applied to data of

the volcano Merapi in Indonesia, collected and provided by the Geophysical Observatory

in Fürstenfeldbruck. In this data set, the specific shape of lagged rain influence on the

occurrence of block and ash flows is examined.

Cloud computing is a paradigm for using IT services with characteristics such as flexible and scalable service usage, on-demand availability, and pay-as-you-go billing. Respective services are called cloud services and their nature usually motivates a differentiation in three layers: Infrastructure as a Service (IaaS) for cloud services offering functionality of hardware resources in a virtualised way, Platform as a Service (PaaS) for services acting as execution platforms, and Software as a Service (SaaS) representing applications provided in a cloud computing way.

Any of these services is offered with the illusion of unlimited scalability. The infinity gained by this illusion implies the need for some kind of regulation mechanism to manage sup- ply and demand. Today’s static pricing mechanisms are limited in their capabilities to adapt to dynamic characteristics of cloud environments such as changing workloads. The solution is a dy- namic pricing approch compareable to today’s exchanges. This requires comparability of cloud services and the need of standardised access to avoid vendor lock-in. To achieve comparability, a classification for cloud services is introcuced, where classes of cloud services representing tradable goods are expressed by the minimum requirements for a certain class.

The main result of this work is a framework for exchange-based trading of cloud com- puting commodities, which is composed of four core components derived from existing ex- change market places. An exchange component takes care of accepting orders from buyers and sellers and determines the price for the goods. A clearing component is responsible for the fi- nancial closing of a trade. The settlement component takes care of the delivery of the cloud service. A rating component monitors cloud services and logs service level agreement breaches to calculate provider ratings, especially for reliability, which is an important factor in cloud computing.

The framework establishes a new basis for using cloud services and more advanced business models. Additionally, an overview of selected economic aspects including ideas for derivative financial instruments like futures, options, insurances, and more complex ones is pro- vided. A first version of the framework is currently being implemented and in use at Deutsche Bo ̈rse Cloud Exchange AG.

Crowdsourcing denotes the transfer of work commonly carried out by single humans to a large group of people. Nowadays, crowdsourcing is employed for many purposes, like people contributing their knowledge to Wikipedia, researchers predicting diseases from data on Twitter, or players solving protein folding problems in games. Still, there are areas for which the application of crowdsourcing has not yet been investigated thoroughly. This thesis examines crowdsourcing for two such areas: for empirical research in sciences oriented on humans -focusing on linguistic field research- and for e-learning.

Sciences oriented on humans -like linguistics, sociology, or art history- depend on empirical research. For example, in traditional linguistic field research researchers ask questions and fill in forms. Such methods are time-consuming, costly, and not free of biases. This thesis proposes the application of crowdsourcing techniques to overcome these disadvantages and to support empirical research in getting more efficient. Therefore, the concept of a generic market for trading with symbolic goods and speculating on their characteristics in a playful manner, called Agora is introduced. Agora aims to be an "operating system" for social media applications gathering data. Furthermore, the Web-based crowdsourcing platform metropolitalia has been established for hosting two social media applications based upon Agora: Mercato Linguistico and Poker Parole. These applications have been conceived as part of this thesis for gathering complementary data and meta-data on Italian language varieties. Mercato Linguistico incites players to express their own knowledge or beliefs, Poker Parole incites players to make conjectures on the contributions of others. Thereby the primary meta-data collected with Mercato Linguistico are enriched with secondary, reflexive meta-data from Poker Parole, which are needed for studies on the perception of languages. An evaluation of the data gathered on metropolitalia exhibits the viability of the market-based approach of Agora and highlights its strengths.

E-learning is concerned with the use of digital technology for learning, nowadays especially via the Internet. This thesis investigates how e-learning applications can support students with association-based learning and lecturers with teaching. For that, a game-like e-learning tool named Termina is proposed in this thesis. From the data collected with Termina association maps are constructed. An association map is a simplified version of a concept map, in which concepts are represented as rectangles and relationships between concepts as links. They constitute an abstract comprehension of a topic. Students profit from the association maps' availability, learn from other participating students, and can track their own learning progress. Lecturers gain insights into the knowledge and into potential misunderstandings of their students. An evaluation of Termina and the collected data along a university course exhibits Termina's usefulness for both students and lecturers.

The main contributions of this thesis are (1) a literature review over collective intelligence, crowdsourcing, and related fields, (2) a model of a generic market for gathering data for empirical research efficiently, (3) two applications based on this model and results of an evaluation of the data gathered with them, (4) the game-like e-learning tool Termina together with insights from its evaluation, and (5) a generic software architecture for all aforementioned applications.

Recent technological advances have made it possible to measure various parameters of biological processes in a genome-wide manner. While traditional molecular biology focusses on individual processes using targeted experiments (reductionistic approach), the field of systems biology utilizes high-throughput experiments to determine the state of a complete system such as a cell at once (holistic approach). Systems biology is not only carried out in wet-lab, but for the most part also requires tailored computational methods. High-throughput experiments are able to produce massive amounts of data, that are often too complex for a human to comprehend directly, that are affected by substantial noise, i.e. random measurement variation, and that are often subject to considerable bias, i.e. systematic deviations of the measurement from the truth. Thus, computer science and statistical methods are necessary for a proper analysis of raw data from such large-scale experiments.

The goal of systems biology is to understand a whole system such as a cell in a quantitative manner. Thus, the computational part does not end with analyzing raw data but also involves visualization, statistical analyses, integration and interpretation. One example for these four computational tasks is as follows: Processes in biological systems are often modeled as networks, for instance, gene regulatory networks (GRNs) that represent the interactions of transcription factors (TFs) and their target genes. Experiments can provide both, the identity and wiring of all constituent parts of the network as well as parameters that allow to describe the processes in the system in a quantative manner. A network provides a straight-forward way to visualize the state and processes of a whole system, its statistical analysis can reveal interesting properties of biological systems, it is able to integrate several datasets from various experiments and simulations of the network can aid to interpret the data.

In recent years, microRNAs emerged as important contributors to gene regulation in eukaryotes, breaking the traditional dogma of molecular biology, where DNA is transcribed to RNA which is subsequently translated into proteins. MicroRNAs are small RNAs that are not translated but functional as RNAs: They are able to target specific messenger RNAs (mRNA) and typically lead to their downregulation. Thus, in addition to TFs, microRNAs also play important roles in GRNs. Interestingly, not only animal genomes including the human genome encode microRNAs, but microRNAs are also encoded by several pathogens such as viruses.

In this work I developed several computational systems biology methods and applied them to high-throughout experimental data in the context of a project about herpes viral microRNAs. Three methods, ALPS, PARma and REA, are designed for the analysis of certain types of raw data, namely short RNA-seq, PAR-CLIP and RIP-Chip data, respectively. All of theses experiments are widely used and my methods are publicly available on the internet and can be utilized by the research community to analyze new datasets. For these methods I developed non-trivial statistical methods (e.g. the EM algorithm kmerExplain in PARma) and implemented and adapted algorithms from traditional computer science and bioinformatics (e.g. alignment of pattern matrices in ALPS).

I applied these novel methods to data measured by our cooperation partners in the herpes virus project. I.a., I discovered and investigated an important aspect of microRNA-mediated regulation: MicroRNAs recognize their targets in a context-dependent manner. The widespread impact of context on regulation is widely accepted for transcriptional regulation, and only few examples are known for microRNA-mediated regulation. By integrating various herpes-related datasets, I could show that context-dependency is not restricted to few examples but is a widespread feature in post-transcriptional regulation mediated by mi