As new sensors are added to VR headsets, more data can be collected. This introduces a new potential threat to user privacy. We focused on the feasibility of extracting personal information from eye-tracking. To achieve this, we designed a preliminary user study focusing on the pupil response to audio stimuli. We used a variation of machine learning models to test the collected data to determine the feasibility of obtaining information such as the age or gender of the participant. Several of the experiments show promise for obtaining this information. We were able to extract with reasonable certainty whether caffeine was consumed and the gender of the participant. This demonstrates the unknown threat that embedded sensors pose to users. A further studies are planned to verify the results.
Machine learning models for timeseries have always been a special topic of interest due to their unique data structure. Recently, the introduction of attention improved the capabilities of recurrent neural networks and transformers with respect to their learning tasks such as machine translation. However, these models are usually subsymbolic architectures, making their inner working hard to interpret without comprehensive tools. In contrast, interpretable models such learning vector quantization are more transparent in the ability to interpret their decision process. This thesis tries to merge attention as a machine learning function with learning vector quantization to better handle timeseries data. A design on such a model is proposed and tested with a dataset used in connection with the attention based transformers. Although the proposed model did not yield the expected results, this work outlines improvements for further research on this approach.
In this paper, we conduct experiments to optimize the learning rates for the Generalized Learning Vector Quantization (GLVQ) model. Our approach leverages insights from cog- nitive science rooted in the profound intricacies of human thinking. Recognizing that human-like thinking has propelled humankind to its current state, we explore the applica- bility of cognitive science principles in enhancing machine learning. Prior research has demonstrated promising results when applying learning rate methods inspired by cognitive science to Learning Vector Quantization (LVQ) models. In this study, we extend this approach to GLVQ models. Specifically, we examine five distinct cognitive science-inspired GLVQ variants: Conditional Probability (CP), Dual Factor Heuristic (DFH), Middle Symmetry (MS), Loose Symmetry (LS), and Loose Symme- try with Rarity (LSR). Our experiments involve a comprehensive analysis of the performance of these cogni- tive science-derived learning rate techniques across various datasets, aiming to identify optimal settings and variants of cognitive science GLVQ model training. Through this research, we seek to unlock new avenues for enhancing the learning process in machine learning models by drawing inspiration from the rich complexities of human cognition. Keywords: machine learning, GLVQ, cognitive science, cognitive bias, learning rate op- timization, optimizers, human-like learning, Conditional Probability (CP), Dual Factor Heuristic (DFH), Middle Symmetry (MS), Loose Symmetry (LS), Loose Symmetry with Rarity (LSR).
Analysis of Continuous Learning Strategies at the Example of Replay-Based Text Classification
(2023)
Continuous learning is a research field that has significantly boosted in recent years due to highly complex machine and deep learning models. Whereas static models need to be retrained entirely from scratch when new data get available, continuous models progressively adapt to new data saving computational resources. In this context, this work analyzes parameters impacting replay-based continuous learning approaches at the example of a data-incremental text classification task using an MLP and LSTM. Generally, it was found that replay improves the results compared to naive approaches but achieves not the performance of a static model. Mainly, the performances increased with more replayed examples, and the number of training iterations has a significant influence as it can partly control the stability-plasticity-trade-off. In contrast, the impact of balancing the buffer and the strategy to select examples to store in the replay buffer were found to have a minor impact on the results in the present case.
Recently a deep neural network architecture designed to work on graph- structured data have been capturing notice as well as getting implemented in various domains and application. However, learning representation (feature embedding) from graphical data picking pace in research and constructing graph(s) from dataset remains a challenge. The ability to map the data to lower dimensions further makes the task easier while providing comfort in applying many operations. Graph neural network (GNN) is one of the novel neural network models that is catching attention as it is outperforming in various applications like recommender systems, social networks, chemical synthesis, and many more. This thesis discusses a unique approach for a fundamental task on graphs; node classification. The feature embedding for a node is aggregated by applying a Recurrent neural network (RNN), then a GNN model is trained to classify a node with the help of aggregated features and Q learning supports in optimizing the shape of neural networks. This thesis starts with the working principles of the Feedforward neural network, recurrent units like simple RNN, Long short-term memory (LSTM), and Gated recurrent unit (GRU), followed by concepts of Reinforcement learning (RL) and the Q learning algorithm. An overview of the fundamentals of graphs, followed by the GNN architecture and workflow, is discussed subsequently. Some basic GNN models are discussed in brief later before it approaches the technical implementation details, the output of the model, and a comparison with a few other models such as GraphSage and Graph attention network (GAN).
In vielen Einsatzbereichen sind digitale Nachbildungen realer Gebäude von großer Wichtigkeit. Die Erstellung dieser Nachbildungen erfordert bei älteren bzw. historischen Gebäuden allerdings meist erheblichen Vermessungs- und Nachbearbeitungsaufwand mit großem Personal- und Zeitbedarf. Häufig wurde ein Gebäude stilistisch an die jeweilige Zeit angepasst, sodass einzelne Zustände nur mit historischem Bildmaterial reproduzierbar sind.
Am Beispiel mehrerer ausgewählter, aktuell existierender Gebäude der Stadt Mittweida sind realitätsnahe, digitale und veränderbare Modelle mittels eines möglichst automatisierten Workflows erstellt.
Die äußere Erscheinung dieser Modelle kann mit dem entwickelten System automatisiert an andere Stile anpasst werden, welche durch z.B. historisches Bildmaterials von Gebäuden vorgegeben sind. Aufgrund der vielfältigen Einsatzbereiche und weiten Verbreitung finden hierfür Verfahren der Photogrammetrie für die Erstellung und neuronale Netze für die Stilanpassung Anwendung, welche auf handelsüblicher Hardware eingesetzt werden können. Eine Evaluierung erfolgte durch bildlichen Vergleich der stilangepassten Modelle mit dem zugehörigen Bildmaterial.
Diese Masterarbeit zeigt einen Ansatz zur Vorhersage von Zugverspätungen mit Hilfe von Supervised Learning. Dazu werden Modelle mit verschiedenen Algorithmen getestet und miteinander verglichen. Außerdem wird gezeigt, wie das entwickelte Vorhersagemodell in eine Blockchain-Anwendung integriert werden kann.
This thesis investigates the efficacy of four machine learning algorithms, namely linear regression, decision tree, random forest and neural network in the task of lead scoring. Specifically, the study evaluates the performance of these algorithms using datasets without sampling and with random under-sampling and over-sampling using SMOTE. The performance of each algorithm is measure using various performance metrics, including accuracy, AUC-ROC, specificity, sensitivity, precision, recall, F1 score, and G-mean. The results indicate that models trained on the dataset without sampling achieved higher accuracy than those trained on the dataset with either random under-sampling or random over-sampling using SMOTE. However, the neural network demonstrated remarkable results on each dataset compared to the other algorithms. These findings provide valuable insights into the effectiveness of machine learning algorithms for lead scoring tasks, particularly when using different sampling techniques. The findings of this study can aid lead management practices in selecting the most suitable algorithm and sampling technique for their needs. Furthermore, the study contributes to the literature by providing a comprehensive evaluation of the performance of machine learning algorithms for lead scoring tasks. This thesis has practical implications for businesses looking to improve their lead management practices, and future research could extend the analysis to other machine learning algorithms or more extensive datasets.
Das Ziel dieser Masterarbeit ist die Evaluierung des Realtime Multi-Person 2D Pose Estimation Frameworks OpenPose. Dazu wird die Forschungsfrage gestellt, bis zu welcher Pixelgröße ein Mensch allgemein von dem System mit einer Sicherheit von über 50% richtig detektiert und dargestellt wird. Um die Forschungsfrage zu beantworten ist eine Studie mit sieben Probanden durchgeführt wurden. Aus der Datenerhebung geht hervor, dass der gesuchte Confidence Value zwischen 110px und 150px Körpergröße in von Menschen digitalen Bildern erreicht wird.
In machine learning, Learning Vector Quantization (LVQ) is well known as supervised vector quantization. LVQ has been studied to generate optimal reference vectors because of its simple and fast learning algorithm [2]. In many tasks of classification, different variants are considered while training a model and a consideration of variants of large margin in LVQ helps to get significant
results [20]. Large margin LVQ (LMLVQ) is to maximize the distance between decision hyperplane and data points. In this thesis, a comparison of different variants of Generalized Learning Vector Quantization (GLVQ) and Large margin in LVQ is proposed along with visualization, implementation and experimental results.
There are multiple ways to gain information about an individual and its health status, but an increasingly popular field in medicine has become the analysis of human breath, which carries a lot of information about metabolic processes within the individuals body. The information in exhaled breath consists of volatile (organic) compounds (VOCs). These VOCs are products of metabolic processes within the individuals body, thus might be an indicator for diseases disturbing those processes. The compounds are to be detected by mass-spectrometric (MS) or ion-mobility spectrometric (IMS) techniques, making the analysis of these compounds not only bounded to exhaled breath. The resulting data is spectral data, capturing concentrations of the VOCs indirectly through intensities. However, a number of about 3000 VOCs [1] could already be determined in human exhaled breath. The number of research paper about VOC-analysis and detection had risen nearly constantly over the last decade 1. Furthermore, the technique to identify VOCs could also be used to capture biomarker from alien species within the individuals body. Extracting VOCs from an individual can be done by non- or minimal invasive techniques. However, the manual identification of VOCs and biomarkers related to a certain disease or infection is not feasible due to the complexity of the sample and often unknown metabolic products, thus automized techniques are needed. [1–4] To establish breath analysis as a diagnosis tool, machine learning methodes could be used. Machine learning has become a popular and common technique when dealing with medical data, due to the rapid analysis. Taking this advantage, breath analysis using machine learning could become the model of choice for diagnosis, keeping in mind that conventional methodes are laboratory based and thus when trying detect bacterial infection need sometimes several days to identify the organism. [5]
Drought is one of the most common and dangerous threats plants have to face, costing the global agricultural sector billions of dollars every year and leading to the loss of tons of harvest. Until people drastically reduce their consumption of animal products or cellular agriculture comes of age, more and more crops will need to be produced to sustain the ever growing human population. Even then, as more areas on earth are becoming prone to drought due to climate change, we may still have to find or breed plant varieties more suitable to grow and prosper in these changing environments.
Plants respond to drought stress with a complex interplay of hormones, transcription factors, and many other functional or regulatory proteins and mapping out this web of agents is no trivial task. In the last two to three decades or so, machine learning has become immensely popular and is increasingly used to find patterns in situations that are too complex for the human mind to overlook. Even though much of the hype is focused on the latest developments in deep learning, relatively simple methods often yield superior results, especially when data is limited and expensive to gather.
This Master Thesis, conducted at the IPK in Gatersleben, develops an approach for shedding light on the phenotypic and transcriptomic processes that occur when a plant is subjected to stress. It centers around a random forest feature selection algorithm and although it is used here to illuminate drought stress response in Arabidopsis thaliana, it can be applied to all kinds of stresses in all kinds of plants.
Genetic sequence variations at the level of gene promoters influence the binding of transcription factors. In plants, this often leads to differential gene expression across natural accessions and crop cultivars. Some of these differences are propagated through molecular networks and lead to macroscopic phenotypes. However, the link between promoter sequence variation and the variation of its activity is not yet well understood. In this project, we use the power of deep learning in 728 genotypes of Arabidopsis thaliana to shed light on some aspects of that link. Convolutional neural networks were successfully implemented to predict the likelihood of a gene being expressed from its promoter sequence. These networks were also capable of highlighting known and putative new sequence motifs causal for the expression of genes. We tested our algorithms in various scenarios, including single and multiple point mutations, as well as indels on synthetic and real promoter sequences and the respective performance characteristics of the algorithm have been estimated. Finally, we showed that the decision boundary to classify genes as expressed and non-expressed depends on the sensitivity of the transcriptome profiling assay and changing it has an impact on the algorithm’s performance.
Data streams change their statistical behaviour over the time. These changes can occur gradually or abruptly with unforeseen reasons, which may effect the expected outcome. Thus it is important to detect concept drift as soon as it occurs. In this thesis we chose distance based methodology to detect presence of concept drift in the data streams. We used generalized learning vector quantization(GLVQ) and generalized matrix learning vector quantization( GMLVQ) classifiers for distance calculation between prototypes and data points. Chi-square and Kolmogorov–Smirnov tests are used to compare the distance distributions of test and train data sets to indicate the drift presence.
Sequences are an important data structure in molecular biology, but unfortunately it is difficult for most machine learning algorithms to handle them, as they rely on vectorial data. Recent approaches include methods that rely on proximity data, such as median and relational Learning Vector Quantization. However, many of them are limited in the size of the data they are able to handle. A standard method to generate vectorial features for sequence data does not exist yet. Consequently, a way to make sequence data accessible to preferably interpretable machine learning algorithms needs to be found. This thesis will therefore investigate a new approach called the Sensor Response Principle, which is being adapted to protein sequences. Accordingly, sequence similarity is measured via pairwise sequence alignments with different sequence alignment algorithms and various substitution matrices. The measurements are then used as input for learning with the Generalized Learning Vector Quantization algorithm. A special focus lies on sequence length variability as it is suspected to affect the sequence alignment score and therefore the discriminative quality of the generated feature vectors. Specific datasets were generated from the Pfam protein family database to address this question. Further, the impact of the number of references and choice of substitution matrices is examined.
In this thesis, we focus on using machine learning to automate manual or rule-based processes for the deduplication task of the data integration process in an enterprise customer experience program. We study the underlying theoretical foundations of the most widely used machine learning algorithms, including logistic regression, random forests, extreme gradient boosting trees, support vector machines, and generalized matrix learning vector quantization. We then apply those algorithms to a real, private data set and use standard evaluation metrics for classification, such as confusion matrix, precision, and recall, area under the precision-recall curve, and area under the Receiver Operating Characteristic curve to compare their performances and results.
Differentiation is ubiquitous in the field of mathematics and especially in the field of Machine learning for calculations in gradient-based models. Calculating gradients might be complex and require handling multiple variables. Supervised Learning Vector Quantization models, which are used for classification tasks, also use the Stochastic Gradient Descent method for optimizing their cost functions. There are various methods to calculate these gradients or derivatives, namely Manual Differentiation, Numeric Differentiation, Symbolic Differentiation, and Automatic Differentiation. In this thesis, we evaluate each of the methods mentioned earlier for calculating derivatives and also compare the use of these methods for the variants of Generalized Learning Vector Quantization algorithms.
Financial fraud for banks can be a reason for huge monetary losses. Studies have shown that, if not mitigated, financial fraud can lead to bankruptcy for big financial institutions and even insolvency for individuals. Credit card fraud is a type of financial fraud that is ever growing. In the future, these numbers are expected to increase exponentially and that’s why a lot of researchers are focusing on machine learning techniques for detecting frauds. This task, however, is not a simple task. There are mainly two reasons
• varying behaviour in committing fraud
• high level of imbalance in the dataset (the majority of normal or genuine cases largely outnumbers the number of fraudulent cases)
A predictive model usually tends to be biased towards the majority of samples, in an unbalanced dataset, when this dataset is provided as an input to a predictive model.
In this Thesis this problem is tackled by implementing a data-level approach where different resampling methods such as undersampling, oversampling, and hybrid strategies along with bagging and boosting algorithmic approaches have been applied to a highly skewed dataset with 492 idetified frauds out of 284,807 transactions.
Predictive modelling algorithms like Logistic Regression, Random Forest, and XGBoost have been implemented along with different resampling techniques to predict fraudulent transactions.
The performance of the predictive models was evaluated based on Receiver Operating CharacteristicArea under the curve (AUC-ROC), Precision Recall Area under the Curve (AUC-PR), Precision, Recall, F1 score metrics.
Embeddings for Product Data
(2022)
The E-commerce industry has grown exponentially in the last decade, with giants like Amazon, eBay, Aliexpress, and Walmart selling billions of products. Machine learning techniques can be used within the e-commerce domain to improve the overall customer journey on a platform and increase sales. Product data, in specific, can be used for various applications, such as product similarity, clustering, recommendation, and price estimation. For data from these products to be used for such applications, we have to perform feature engineering. The idea is to transform these products into feature vectors before training a machine learning model on them. In this thesis, we propose an approach to create representations for heterogeneous product data from Unite’s platform in the form of structured tabular records. These tables consist of attributes having different information ranging from product-ids to long descriptions. Our model combines popular deep learning approaches used in natural language processing to create numerical representations, which contain mostly non-zeros elements in an array or matrix called as dense representation for all products. To evaluate the quality of these feature vectors, we validate how well the similarities between products are captured by these dense representations. The evaluations are further divided into two categories. The first category directly compares the similarities between individual products. On the other hand, the second category uses these dense vectors in any of the above- mentioned applications as inputs. It then evaluates the quality of these dense representation vectors based on the accuracy or performance of the defined application. As result, we explain the impact of different steps within our model on the quality of these learned representations.