@phdthesis{Stojanovic2021,
  author    = {Stojanovic, Vladeta},
  title     = {Digital twins for indoor built environments},
  doi       = {10.25932/publishup-50913},
  url       = {http://nbn-resolving.de/urn:nbn:de:kobv:517-opus4-509134},
  school      = {Universit{\"a}t Potsdam},
  pages     = {xxiii, 181},
  year      = {2021},
  abstract  = {One of the key challenges in modern Facility Management (FM) is to digitally reflect the current state of the built environment, referred to as-is or as-built versus as-designed representation. While the use of Building Information Modeling (BIM) can address the issue of digital representation, the generation and maintenance of BIM data requires a considerable amount of manual work and domain expertise. Another key challenge is being able to monitor the current state of the built environment, which is used to provide feedback and enhance decision making. The need for an integrated solution for all data associated with the operational life cycle of a building is becoming more pronounced as practices from Industry 4.0 are currently being evaluated and adopted for FM use. This research presents an approach for digital representation of indoor environments in their current state within the life cycle of a given building. Such an approach requires the fusion of various sources of digital data. The key to solving such a complex issue of digital data integration, processing and representation is with the use of a Digital Twin (DT). A DT is a digital duplicate of the physical environment, states, and processes. A DT fuses as-designed and as-built digital representations of built environment with as-is data, typically in the form of floorplans, point clouds and BIMs, with additional information layers pertaining to the current and predicted states of an indoor environment or a complete building (e.g., sensor data). The design, implementation and initial testing of prototypical DT software services for indoor environments is presented and described. These DT software services are implemented within a service-oriented paradigm, and their feasibility is presented through functioning and tested key software components within prototypical Service-Oriented System (SOS) implementations. The main outcome of this research shows that key data related to the built environment can be semantically enriched and combined to enable digital representations of indoor environments, based on the concept of a DT. Furthermore, the outcomes of this research show that digital data, related to FM and Architecture, Construction, Engineering, Owner and Occupant (AECOO) activity, can be combined, analyzed and visualized in real-time using a service-oriented approach. This has great potential to benefit decision making related to Operation and Maintenance (O\&M) procedures within the scope of the post-construction life cycle stages of typical office buildings.},
  language  = {en}
}
@phdthesis{Koumarelas2020,
  author    = {Koumarelas, Ioannis},
  title     = {Data preparation and domain-agnostic duplicate detection},
  doi       = {10.25932/publishup-48913},
  url       = {http://nbn-resolving.de/urn:nbn:de:kobv:517-opus4-489131},
  school      = {Universit{\"a}t Potsdam},
  pages     = {x, 97},
  year      = {2020},
  abstract  = {Successfully completing any data science project demands careful consideration across its whole process. Although the focus is often put on later phases of the process, in practice, experts spend more time in earlier phases, preparing data, to make them consistent with the systems' requirements or to improve their models' accuracies. Duplicate detection is typically applied during the data cleaning phase, which is dedicated to removing data inconsistencies and improving the overall quality and usability of data. While data cleaning involves a plethora of approaches to perform specific operations, such as schema alignment and data normalization, the task of detecting and removing duplicate records is particularly challenging. Duplicates arise when multiple records representing the same entities exist in a database. Due to numerous reasons, spanning from simple typographical errors to different schemas and formats of integrated databases. Keeping a database free of duplicates is crucial for most use-cases, as their existence causes false negatives and false positives when matching queries against it. These two data quality issues have negative implications for tasks, such as hotel booking, where users may erroneously select a wrong hotel, or parcel delivery, where a parcel can get delivered to the wrong address. Identifying the variety of possible data issues to eliminate duplicates demands sophisticated approaches. While research in duplicate detection is well-established and covers different aspects of both efficiency and effectiveness, our work in this thesis focuses on the latter. We propose novel approaches to improve data quality before duplicate detection takes place and apply the latter in datasets even when prior labeling is not available. Our experiments show that improving data quality upfront can increase duplicate classification results by up to 19\%. To this end, we propose two novel pipelines that select and apply generic as well as address-specific data preparation steps with the purpose of maximizing the success of duplicate detection. Generic data preparation, such as the removal of special characters, can be applied to any relation with alphanumeric attributes. When applied, data preparation steps are selected only for attributes where there are positive effects on pair similarities, which indirectly affect classification, or on classification directly. Our work on addresses is twofold; first, we consider more domain-specific approaches to improve the quality of values, and, second, we experiment with known and modified versions of similarity measures to select the most appropriate per address attribute, e.g., city or country. To facilitate duplicate detection in applications where gold standard annotations are not available and obtaining them is not possible or too expensive, we propose MDedup. MDedup is a novel, rule-based, and fully automatic duplicate detection approach that is based on matching dependencies. These dependencies can be used to detect duplicates and can be discovered using state-of-the-art algorithms efficiently and without any prior labeling. MDedup uses two pipelines to first train on datasets with known labels, learning to identify useful matching dependencies, and then be applied on unseen datasets, regardless of any existing gold standard. Finally, our work is accompanied by open source code to enable repeatability of our research results and application of our approaches to other datasets.},
  language  = {en}
}
@phdthesis{Risch2020,
  author    = {Risch, Julian},
  title     = {Reader comment analysis on online news platforms},
  doi       = {10.25932/publishup-48922},
  url       = {http://nbn-resolving.de/urn:nbn:de:kobv:517-opus4-489222},
  school      = {Universit{\"a}t Potsdam},
  pages     = {xi, 135},
  year      = {2020},
  abstract  = {Comment sections of online news platforms are an essential space to express opinions and discuss political topics. However, the misuse by spammers, haters, and trolls raises doubts about whether the benefits justify the costs of the time-consuming content moderation. As a consequence, many platforms limited or even shut down comment sections completely. In this thesis, we present deep learning approaches for comment classification, recommendation, and prediction to foster respectful and engaging online discussions. The main focus is on two kinds of comments: toxic comments, which make readers leave a discussion, and engaging comments, which make readers join a discussion. First, we discourage and remove toxic comments, e.g., insults or threats. To this end, we present a semi-automatic comment moderation process, which is based on fine-grained text classification models and supports moderators. Our experiments demonstrate that data augmentation, transfer learning, and ensemble learning allow training robust classifiers even on small datasets. To establish trust in the machine-learned models, we reveal which input features are decisive for their output with attribution-based explanation methods. Second, we encourage and highlight engaging comments, e.g., serious questions or factual statements. We automatically identify the most engaging comments, so that readers need not scroll through thousands of comments to find them. The model training process builds on upvotes and replies as a measure of reader engagement. We also identify comments that address the article authors or are otherwise relevant to them to support interactions between journalists and their readership. Taking into account the readers' interests, we further provide personalized recommendations of discussions that align with their favored topics or involve frequent co-commenters. Our models outperform multiple baselines and recent related work in experiments on comment datasets from different platforms.},
  language  = {en}
}
@phdthesis{Zhelavskaya2020,
  author    = {Zhelavskaya, Irina},
  title     = {Modeling of the Plasmasphere Dynamics},
  doi       = {10.25932/publishup-48243},
  url       = {http://nbn-resolving.de/urn:nbn:de:kobv:517-opus4-482433},
  school      = {Universit{\"a}t Potsdam},
  pages     = {xlii, 256},
  year      = {2020},
  abstract  = {The plasmasphere is a dynamic region of cold, dense plasma surrounding the Earth. Its shape and size are highly susceptible to variations in solar and geomagnetic conditions. Having an accurate model of plasma density in the plasmasphere is important for GNSS navigation and for predicting hazardous effects of radiation in space on spacecraft. The distribution of cold plasma and its dynamic dependence on solar wind and geomagnetic conditions remain, however, poorly quantified. Existing empirical models of plasma density tend to be oversimplified as they are based on statistical averages over static parameters. Understanding the global dynamics of the plasmasphere using observations from space remains a challenge, as existing density measurements are sparse and limited to locations where satellites can provide in-situ observations. In this dissertation, we demonstrate how such sparse electron density measurements can be used to reconstruct the global electron density distribution in the plasmasphere and capture its dynamic dependence on solar wind and geomagnetic conditions. First, we develop an automated algorithm to determine the electron density from in-situ measurements of the electric field on the Van Allen Probes spacecraft. In particular, we design a neural network to infer the upper hybrid resonance frequency from the dynamic spectrograms obtained with the Electric and Magnetic Field Instrument Suite and Integrated Science (EMFISIS) instrumentation suite, which is then used to calculate the electron number density. The developed Neural-network-based Upper hybrid Resonance Determination (NURD) algorithm is applied to more than four years of EMFISIS measurements to produce the publicly available electron density data set. We utilize the obtained electron density data set to develop a new global model of plasma density by employing a neural network-based modeling approach. In addition to the location, the model takes the time history of geomagnetic indices and location as inputs, and produces electron density in the equatorial plane as an output. It is extensively validated using in-situ density measurements from the Van Allen Probes mission, and also by comparing the predicted global evolution of the plasmasphere with the global IMAGE EUV images of He+ distribution. The model successfully reproduces erosion of the plasmasphere on the night side as well as plume formation and evolution, and agrees well with data. The performance of neural networks strongly depends on the availability of training data, which is limited during intervals of high geomagnetic activity. In order to provide reliable density predictions during such intervals, we can employ physics-based modeling. We develop a new approach for optimally combining the neural network- and physics-based models of the plasmasphere by means of data assimilation. The developed approach utilizes advantages of both neural network- and physics-based modeling and produces reliable global plasma density reconstructions for quiet, disturbed, and extreme geomagnetic conditions. Finally, we extend the developed machine learning-based tools and apply them to another important problem in the field of space weather, the prediction of the geomagnetic index Kp. The Kp index is one of the most widely used indicators for space weather alerts and serves as input to various models, such as for the thermosphere, the radiation belts and the plasmasphere. It is therefore crucial to predict the Kp index accurately. Previous work in this area has mostly employed artificial neural networks to nowcast and make short-term predictions of Kp, basing their inferences on the recent history of Kp and solar wind measurements at L1. We analyze how the performance of neural networks compares to other machine learning algorithms for nowcasting and forecasting Kp for up to 12 hours ahead. Additionally, we investigate several machine learning and information theory methods for selecting the optimal inputs to a predictive model of Kp. The developed tools for feature selection can also be applied to other problems in space physics in order to reduce the input dimensionality and identify the most important drivers. Research outlined in this dissertation clearly demonstrates that machine learning tools can be used to develop empirical models from sparse data and also can be used to understand the underlying physical processes. Combining machine learning, physics-based modeling and data assimilation allows us to develop novel methods benefiting from these different approaches.},
  language  = {en}
}
@phdthesis{Dick2016,
  author    = {Dick, Uwe},
  title     = {Discriminative Classification Models for Internet Security},
  url       = {http://nbn-resolving.de/urn:nbn:de:kobv:517-opus4-102593},
  school      = {Universit{\"a}t Potsdam},
  pages     = {x, 57},
  year      = {2016},
  abstract  = {Services that operate over the Internet are under constant threat of being exposed to fraudulent use. Maintaining good user experience for legitimate users often requires the classification of entities as malicious or legitimate in order to initiate countermeasures. As an example, inbound email spam filters decide for spam or non-spam. They can base their decision on both the content of each email as well as on features that summarize prior emails received from the sending server. In general, discriminative classification methods learn to distinguish positive from negative entities. Each decision for a label may be based on features of the entity and related entities. When labels of related entities have strong interdependencies---as can be assumed e.g. for emails being delivered by the same user---classification decisions should not be made independently and dependencies should be modeled in the decision function. This thesis addresses the formulation of discriminative classification problems that are tailored for the specific demands of the following three Internet security applications. Theoretical and algorithmic solutions are devised to protect an email service against flooding of user inboxes, to mitigate abusive usage of outbound email servers, and to protect web servers against distributed denial of service attacks. In the application of filtering an inbound email stream for unsolicited emails, utilizing features that go beyond each individual email's content can be valuable. Information about each sending mail server can be aggregated over time and may help in identifying unwanted emails. However, while this information will be available to the deployed email filter, some parts of the training data that are compiled by third party providers may not contain this information. The missing features have to be estimated at training time in order to learn a classification model. In this thesis an algorithm is derived that learns a decision function that integrates over a distribution of values for each missing entry. The distribution of missing values is a free parameter that is optimized to learn an optimal decision function. The outbound stream of emails of an email service provider can be separated by the customer IDs that ask for delivery. All emails that are sent by the same ID in the same period of time are related, both in content and in label. Hijacked customer accounts may send batches of unsolicited emails to other email providers, which in turn might blacklist the sender's email servers after detection of incoming spam emails. The risk of being blocked from further delivery depends on the rate of outgoing unwanted emails and the duration of high spam sending rates. An optimization problem is developed that minimizes the expected cost for the email provider by learning a decision function that assigns a limit on the sending rate to customers based on the each customer's email stream. Identifying attacking IPs during HTTP-level DDoS attacks allows to block those IPs from further accessing the web servers. DDoS attacks are usually carried out by infected clients that are members of the same botnet and show similar traffic patterns. HTTP-level attacks aim at exhausting one or more resources of the web server infrastructure, such as CPU time. If the joint set of attackers cannot increase resource usage close to the maximum capacity, no effect will be experienced by legitimate users of hosted web sites. However, if the additional load raises the computational burden towards the critical range, user experience will degrade until service may be unavailable altogether. As the loss of missing one attacker depends on block decisions for other attackers---if most other attackers are detected, not blocking one client will likely not be harmful---a structured output model has to be learned. In this thesis an algorithm is developed that learns a structured prediction decoder that searches the space of label assignments, guided by a policy. Each model is evaluated on real-world data and is compared to reference methods. The results show that modeling each classification problem according to the specific demands of the task improves performance over solutions that do not consider the constraints inherent to an application.},
  language  = {en}
}
@phdthesis{Brueckner2012,
  author    = {Br{\"u}ckner, Michael},
  title     = {Prediction games : machine learning in the presence of an adversary},
  publisher = {Universit{\"a}tsverlag Potsdam},
  address   = {Potsdam},
  isbn      = {978-3-86956-203-2},
  url       = {http://nbn-resolving.de/urn:nbn:de:kobv:517-opus-60375},
  school      = {Universit{\"a}t Potsdam},
  pages     = {x, 121},
  year      = {2012},
  abstract  = {In many applications one is faced with the problem of inferring some functional relation between input and output variables from given data. Consider, for instance, the task of email spam filtering where one seeks to find a model which automatically assigns new, previously unseen emails to class spam or non-spam. Building such a predictive model based on observed training inputs (e.g., emails) with corresponding outputs (e.g., spam labels) is a major goal of machine learning. Many learning methods assume that these training data are governed by the same distribution as the test data which the predictive model will be exposed to at application time. That assumption is violated when the test data are generated in response to the presence of a predictive model. This becomes apparent, for instance, in the above example of email spam filtering. Here, email service providers employ spam filters and spam senders engineer campaign templates such as to achieve a high rate of successful deliveries despite any filters. Most of the existing work casts such situations as learning robust models which are unsusceptible against small changes of the data generation process. The models are constructed under the worst-case assumption that these changes are performed such to produce the highest possible adverse effect on the performance of the predictive model. However, this approach is not capable to realistically model the true dependency between the model-building process and the process of generating future data. We therefore establish the concept of prediction games: We model the interaction between a learner, who builds the predictive model, and a data generator, who controls the process of data generation, as an one-shot game. The game-theoretic framework enables us to explicitly model the players' interests, their possible actions, their level of knowledge about each other, and the order at which they decide for an action. We model the players' interests as minimizing their own cost function which both depend on both players' actions. The learner's action is to choose the model parameters and the data generator's action is to perturbate the training data which reflects the modification of the data generation process with respect to the past data. We extensively study three instances of prediction games which differ regarding the order in which the players decide for their action. We first assume that both player choose their actions simultaneously, that is, without the knowledge of their opponent's decision. We identify conditions under which this Nash prediction game has a meaningful solution, that is, a unique Nash equilibrium, and derive algorithms that find the equilibrial prediction model. As a second case, we consider a data generator who is potentially fully informed about the move of the learner. This setting establishes a Stackelberg competition. We derive a relaxed optimization criterion to determine the solution of this game and show that this Stackelberg prediction game generalizes existing prediction models. Finally, we study the setting where the learner observes the data generator's action, that is, the (unlabeled) test data, before building the predictive model. As the test data and the training data may be governed by differing probability distributions, this scenario reduces to learning under covariate shift. We derive a new integrated as well as a two-stage method to account for this data set shift. In case studies on email spam filtering we empirically explore properties of all derived models as well as several existing baseline methods. We show that spam filters resulting from the Nash prediction game as well as the Stackelberg prediction game in the majority of cases outperform other existing baseline methods.},
  language  = {en}
}
@phdthesis{Sawade2012,
  author    = {Sawade, Christoph},
  title     = {Active evaluation of predictive models},
  publisher = {Universit{\"a}tsverlag Potsdam},
  address   = {Potsdam},
  isbn      = {978-3-86956-255-1},
  url       = {http://nbn-resolving.de/urn:nbn:de:kobv:517-opus-65583},
  school      = {Universit{\"a}t Potsdam},
  pages     = {ix, 157},
  year      = {2012},
  abstract  = {The field of machine learning studies algorithms that infer predictive models from data. Predictive models are applicable for many practical tasks such as spam filtering, face and handwritten digit recognition, and personalized product recommendation. In general, they are used to predict a target label for a given data instance. In order to make an informed decision about the deployment of a predictive model, it is crucial to know the model's approximate performance. To evaluate performance, a set of labeled test instances is required that is drawn from the distribution the model will be exposed to at application time. In many practical scenarios, unlabeled test instances are readily available, but the process of labeling them can be a time- and cost-intensive task and may involve a human expert. This thesis addresses the problem of evaluating a given predictive model accurately with minimal labeling effort. We study an active model evaluation process that selects certain instances of the data according to an instrumental sampling distribution and queries their labels. We derive sampling distributions that minimize estimation error with respect to different performance measures such as error rate, mean squared error, and F-measures. An analysis of the distribution that governs the estimator leads to confidence intervals, which indicate how precise the error estimation is. Labeling costs may vary across different instances depending on certain characteristics of the data. For instance, documents differ in their length, comprehensibility, and technical requirements; these attributes affect the time a human labeler needs to judge relevance or to assign topics. To address this, the sampling distribution is extended to incorporate instance-specific costs. We empirically study conditions under which the active evaluation processes are more accurate than a standard estimate that draws equally many instances from the test distribution. We also address the problem of comparing the risks of two predictive models. The standard approach would be to draw instances according to the test distribution, label the selected instances, and apply statistical tests to identify significant differences. Drawing instances according to an instrumental distribution affects the power of a statistical test. We derive a sampling procedure that maximizes test power when used to select instances, and thereby minimizes the likelihood of choosing the inferior model. Furthermore, we investigate the task of comparing several alternative models; the objective of an evaluation could be to rank the models according to the risk that they incur or to identify the model with lowest risk. An experimental study shows that the active procedure leads to higher test power than the standard test in many application domains. Finally, we study the problem of evaluating the performance of ranking functions, which are used for example for web search. In practice, ranking performance is estimated by applying a given ranking model to a representative set of test queries and manually assessing the relevance of all retrieved items for each query. We apply the concepts of active evaluation and active comparison to ranking functions and derive optimal sampling distributions for the commonly used performance measures Discounted Cumulative Gain and Expected Reciprocal Rank. Experiments on web search engine data illustrate significant reductions in labeling costs.},
  language  = {en}
}
@phdthesis{Bickel2008,
  author    = {Bickel, Steffen},
  title     = {Learning under differing training and test distributions},
  url       = {http://nbn-resolving.de/urn:nbn:de:kobv:517-opus-33331},
  school      = {Universit{\"a}t Potsdam},
  year      = {2008},
  abstract  = {One of the main problems in machine learning is to train a predictive model from training data and to make predictions on test data. Most predictive models are constructed under the assumption that the training data is governed by the exact same distribution which the model will later be exposed to. In practice, control over the data collection process is often imperfect. A typical scenario is when labels are collected by questionnaires and one does not have access to the test population. For example, parts of the test population are underrepresented in the survey, out of reach, or do not return the questionnaire. In many applications training data from the test distribution are scarce because they are difficult to obtain or very expensive. Data from auxiliary sources drawn from similar distributions are often cheaply available. This thesis centers around learning under differing training and test distributions and covers several problem settings with different assumptions on the relationship between training and test distributions-including multi-task learning and learning under covariate shift and sample selection bias. Several new models are derived that directly characterize the divergence between training and test distributions, without the intermediate step of estimating training and test distributions separately. The integral part of these models are rescaling weights that match the rescaled or resampled training distribution to the test distribution. Integrated models are studied where only one optimization problem needs to be solved for learning under differing distributions. With a two-step approximation to the integrated models almost any supervised learning algorithm can be adopted to biased training data. In case studies on spam filtering, HIV therapy screening, targeted advertising, and other applications the performance of the new models is compared to state-of-the-art reference methods.},
  language  = {en}
}
@phdthesis{Dornhege2006,
  author    = {Dornhege, Guido},
  title     = {Increasing information transfer rates for brain-computer interfacing},
  url       = {http://nbn-resolving.de/urn:nbn:de:kobv:517-opus-7690},
  school      = {Universit{\"a}t Potsdam},
  year      = {2006},
  abstract  = {The goal of a Brain-Computer Interface (BCI) consists of the development of a unidirectional interface between a human and a computer to allow control of a device only via brain signals. While the BCI systems of almost all other groups require the user to be trained over several weeks or even months, the group of Prof. Dr. Klaus-Robert M{\"u}ller in Berlin and Potsdam, which I belong to, was one of the first research groups in this field which used machine learning techniques on a large scale. The adaptivity of the processing system to the individual brain patterns of the subject confers huge advantages for the user. Thus BCI research is considered a hot topic in machine learning and computer science. It requires interdisciplinary cooperation between disparate fields such as neuroscience, since only by combining machine learning and signal processing techniques based on neurophysiological knowledge will the largest progress be made. In this work I particularly deal with my part of this project, which lies mainly in the area of computer science. I have considered the following three main points: <b>Establishing a performance measure based on information theory:</b> I have critically illuminated the assumptions of Shannon's information transfer rate for application in a BCI context. By establishing suitable coding strategies I was able to show that this theoretical measure approximates quite well to what is practically achieveable. <b>Transfer and development of suitable signal processing and machine learning techniques:</b> One substantial component of my work was to develop several machine learning and signal processing algorithms to improve the efficiency of a BCI. Based on the neurophysiological knowledge that several independent EEG features can be observed for some mental states, I have developed a method for combining different and maybe independent features which improved performance. In some cases the performance of the combination algorithm outperforms the best single performance by more than 50 \%. Furthermore, I have theoretically and practically addressed via the development of suitable algorithms the question of the optimal number of classes which should be used for a BCI. It transpired that with BCI performances reported so far, three or four different mental states are optimal. For another extension I have combined ideas from signal processing with those of machine learning since a high gain can be achieved if the temporal filtering, i.e., the choice of frequency bands, is automatically adapted to each subject individually. <b>Implementation of the Berlin brain computer interface and realization of suitable experiments:</b> Finally a further substantial component of my work was to realize an online BCI system which includes the developed methods, but is also flexible enough to allow the simple realization of new algorithms and ideas. So far, bitrates of up to 40 bits per minute have been achieved with this system by absolutely untrained users which, compared to results of other groups, is highly successful.},
  subject      = {Kybernetik},
  language  = {en}
}
@phdthesis{Scholz2006,
  author    = {Scholz, Matthias},
  title     = {Approaches to analyse and interpret biological profile data},
  url       = {http://nbn-resolving.de/urn:nbn:de:kobv:517-opus-7839},
  school      = {Universit{\"a}t Potsdam},
  year      = {2006},
  abstract  = {Advances in biotechnologies rapidly increase the number of molecules of a cell which can be observed simultaneously. This includes expression levels of thousands or ten-thousands of genes as well as concentration levels of metabolites or proteins. Such Profile data, observed at different times or at different experimental conditions (e.g., heat or dry stress), show how the biological experiment is reflected on the molecular level. This information is helpful to understand the molecular behaviour and to identify molecules or combination of molecules that characterise specific biological condition (e.g., disease). This work shows the potentials of component extraction algorithms to identify the major factors which influenced the observed data. This can be the expected experimental factors such as the time or temperature as well as unexpected factors such as technical artefacts or even unknown biological behaviour. Extracting components means to reduce the very high-dimensional data to a small set of new variables termed components. Each component is a combination of all original variables. The classical approach for that purpose is the principal component analysis (PCA). It is shown that, in contrast to PCA which maximises the variance only, modern approaches such as independent component analysis (ICA) are more suitable for analysing molecular data. The condition of independence between components of ICA fits more naturally our assumption of individual (independent) factors which influence the data. This higher potential of ICA is demonstrated by a crossing experiment of the model plant Arabidopsis thaliana (Thale Cress). The experimental factors could be well identified and, in addition, ICA could even detect a technical artefact. However, in continuously observations such as in time experiments, the data show, in general, a nonlinear distribution. To analyse such nonlinear data, a nonlinear extension of PCA is used. This nonlinear PCA (NLPCA) is based on a neural network algorithm. The algorithm is adapted to be applicable to incomplete molecular data sets. Thus, it provides also the ability to estimate the missing data. The potential of nonlinear PCA to identify nonlinear factors is demonstrated by a cold stress experiment of Arabidopsis thaliana. The results of component analysis can be used to build a molecular network model. Since it includes functional dependencies it is termed functional network. Applied to the cold stress data, it is shown that functional networks are appropriate to visualise biological processes and thereby reveals molecular dynamics.},
  subject      = {Bioinformatik},
  language  = {en}
}