Refine
Has Fulltext
- yes (23)
Document Type
- Master's Thesis (17)
- Bachelor Thesis (6)
Keywords
- Deep Learning (5)
- LSTM (4)
- Transfer Learning (4)
- CNN (2)
- GloVe (2)
- ULMFiT (2)
- ARIMA modelling (1)
- Action Design Research (1)
- Active Portfolio Management (1)
- Algorithmic Trading (1)
Moderne Industrie ist datengetrieben. Neueste Technologien, Produkte und Dienstleistungen können nur so gut sein, wie die zugrundeliegenden Informationen. Dies gilt für nahezu jede Branche. In Finanzsektor, Telekommunikation, Gesundheitswesen oder Logistik bilden Daten die Grundlage für alle wichtigen Geschäftsentscheidungen.
Verminderte Wettbewerbsfähigkeit, Rückgang der Kundenzufriedenheit oder erhöhtes Risiko von behördlichen Strafen sind unweigerliche Konsequenzen schlechter Datenqualität. Zahlreiche Softwarelösungen wurden in den letzten Jahren entwickelt, um dem Problem entgegenzuwirken.
Diese Arbeit untersucht die Datenqualitätsmanagement-Software, die im Microsoft SQL Server 2016 verankert ist und sich aus drei ergänzenden Teilkomponenten zusammensetzt: Data Quality Services (DQS), SQL Server Integration Services (SSIS) und Master Data Services (MDS). Es werden zunächst theoretische Aspekte der Datenqualität sowie die auf dem Markt vorhandenen Typen von Softwarelösungen betrachtet und daraus Anforderungen an eine Data Quality Suite abgeleitet. Im Anschluss wird der Funktionsumfang mithilfe von Testdaten im Hinblick auf die Erfüllung der Anforderungskriterien analysiert und bewertet. Insbesondere sollen Möglichkeiten des Data Profilings und Data Cleansings mit speziellem Fokus auf Duplikaterkennung und Konsolidierung untersucht werden. Ein großer Wert wird dabei auf die praktische Anwendung der getesteten Software gelegt. Alle Funktionen von SQL Server 2016 werden anschließend auf Basis der definierten Evaluationskriterien bewertet.
A Deep Reinforcement Learning Agent Using Multiple Assets Financial Signals for Portfolio Management
(2019)
In this study, we investigated possible applications of reinforcement learning in the area of portfolio management. Aa specific topology of reinforcement learning is choosen to study the feasibility, Deep Deterministic Policy Gradient (DDPG), to train neural networks to perform trade in an environment simulated real world trading. The results show that the DDPG agent is able to learn price pattern and perform profitable trades. In single stock backtest, the DDPG is able to generate an annual return of 7%. While in multiple stocks backtest, DDPG agent can generate an annual return of 12%.
In this study, we investigate the usage of generative adversarial neural networks(GAN) for a sequence labeling task. We applied sequence generative adversarial neural network (SeqGAN) to extract cause-effect and moderator-mediator relations from hypotheses from scientific papers. This research focuses on the structure of SeqGAN and the problems that come with it. However, there are two main problems with GAN’s. Firstly, a vanilla GAN (Goodfellow et. Al. 2014) is designed for generating real-valued, continuous data but we want to label discrete words into tokens. The second problem is that a GAN can only provide a loss for a complete sequence (Yu, Zhang, Wang, & Yu, 2017, p. 1). To address these problems two things are changed. First, the sequence generation process is realized as a decisionmaking process. The sequences are going to be evaluated by the discriminator model. To solve the problem with the discrete data we follow Yu et. Al. and regard the generative model as a policy gradient. To approximate the sequence value a Monte Carlo Search is employed in the generative model. We show the performance of different parameter settings and which tricks improve the results. The python and r scripts are appended in this document. They can also be found in GitHub (https://github.com/clamkewitz/GANCause).
This research paper explores the application of the Universal Language Model Fine-tuning (ULMFiT) technique, a novel deep transfer learning approach in the Natural Language Processing (NLP) field, to the financial statements fraud detection task. Additionally, the artifact investigated a simpler model, represented by a one-dimensional Convolutional Neural Network (CNN). Both methods have been assessed with respect to the training time and predefined evaluation metrics. Overall, ULMFiT turned out to be considerably more computationally expensive to train and achieved an accuracy of 77% with F-measure of 13%, if used with a decision boundary of 0.5, and accuracy of 59% with F-measure of 42%, if calculated with a threshold of 0.2. In contrast, CNN model trained significantly faster and obtained the following metrics: accuracy of 82% with F-measure of 42% for threshold 0.5, and accuracy of 83% with F-measure of 58% for threshold 0.2. As a result, ULMFiT has been outperformed by one-dimensional CNN on all examined metrics. The results are reported by using two different values for decision boundary due to the precision-recall trade-off, depending on the use case. In addition, this thesis investigated the impact of data preprocessing. The findings have shown that removing all numbers and special symbols with supplementary text truncation, limiting the sequence length, had a positive effect on both models, mentioned above.
This thesis examines the effectiveness of the latest Transfer Learning techniques for Natural Language Processing applied to the classification of research methods used in scientific journals in the domain of Information Systems. The task of automated knowledge extraction from academic articles has seen ongoing progress in recent years. However, the combination of transfer and Deep Learning in order to assign research methods to scientific papers has not been addressed in the literature yet. The main contribution of this thesis is, therefore, an artifact that applies cutting-edge Transfer Learning techniques to a Deep Learning model by conducting several experiments and comparing their effectiveness. The prototype considers various ways of fine-tuning that are crucial to retain the knowledge transferred from pretrained models and avoid catastrophic forgetting. Additionally, this work discusses the literature with regard to the task-specific Theory Ontology Learning and the method-specific state of the art in Transfer Learning for Natural Language Processing. As a result, the artifact surpassed the performance of previously developed models for research method extraction, presented in the literature, without applying any custom feature engineering and only using around a thousand of labeled observations.
This thesis analyzes whether financial news articles can predict the stock prices of companies and the S&P 500 Index. The selected companies for the analysis are Facebook Inc., Apple Inc., Microsoft Corp., Google (Alphabet Inc.), and Amazon.com Inc. This thesis evaluates the predictive power of financial news by comparing the sentiment of financial news articles published on a business day to
the corresponding closing value of stock prices and the S&P 500 Index. For the analysis, financial news from various resources is used for a time period between January 2018 and May 2018. The sentiment of the news articles is analyzed using a lexicon-based approach. Then, regression models are used to predict the stock market. The models used for forecasting the stock prices and the S&P 500 Index
are Auto-Regressive Integrated Moving Average (ARIMA), Support Vector
Regression (SVR), and Linear Regression (LR). The results of the research in overall indicate that ARIMA performs better for predicting the stock prices and the S&P 500 Index.
Hypothesis Extraction from Academic Papers Using Neural Networks for Ontology Theory Learning
(2019)
In this study, we investigated a new use case of deep learning. We applied deep learning to extract causes and effects from the hypotheses of the scientific papers. The research presents a variety of RNN models, including RNN models with CRF layer for labelling the sequences. We used such models as Bi-LSTM, LSTM, SimpleRNN and GRU. The experiments were conducted with GloVe vector representation and character level vector representation of words. Moreover, along with RNN models, we evaluated various hyperparameters and model setups to achieve the highest performance scores. In the end, we obtained promising results and shared our thoughts on the future prospects of the following studies.
An increasing amount of research papers published every year makes it more and more difficult for researchers to find papers that address the same or a similar topic as their own current work. Electronic online databases provide a good starting point when trying to get an overview of the papers published regarding a certain topic. The results presented by the online libraries, however, vary greatly in terms of relevance to the intended search query. In addition, most online libraries provide thousands of search results, of which a huge amount is not relevant for the user at all. There are recent approaches to increase the quality of the search results when looking for relevant research papers online. While most current search engines focus mainly on the title of the papers as well as a few tags, these approaches try to use other information found in the papers as their input.
This master’s thesis focuses on undirected graphs and presents a first prototype for a software that translates images of graphs into a metalanguage. Since many scientific publications contain graphs, they could be a possible future input for search algorithms. For this purpose, an image recognition model consisting of a convolutional neural network and two long short-term memory networks was developed in collaboration with Badr Bih. The model itself is based on a similar software prototype called pix2code developed by Tony Beltramelli in 2017. For computational reasons, we had to limit our graphs to a maximum of four nodes and trained our model on them. The results of the model are promising, but it will have to be trained again on graphs with more vertices to make a definitive evaluation of the approach possible.
Process mining is on the rise. It consists of methods, techniques, and tools to discover, monitor and improve processes by extracting knowledge from event logs of information systems. In combination with the rapidly growing availability of event data (Big Data), challenges
arise in the field of data quality. As data quality is the utmost success factor for process mining, there is an urgent need to get familiar with the topic and leverage possibilities to improve the data quality of event logs.
This thesis contributes to the improvement of event log data quality. It outlines criteria and issues for both, data quality in general and for process mining based on a literature review. The aim is to offer an overview of existing challenges and how to approach them with a preliminary framework, which consists of standard- ized BPMN processes. In addition, design science with experimental validation was applied to master the data quality challenges of data noise and
partial-incomplete traces with common tools like ProM and R. The approach was evaluated through real-life and artificial data.
The results reveal a large variety of data quality challenges for process mining and a need to raise awareness for this topic. Due to the early stage of process mining research, the number of methods, techniques, and tools to improve data quality is limited. Nevertheless, possibilities exist to improve the data quality of event logs. In particular, a repairing technique for data noise was positively evaluated.
The author recommends further specification of the framework and development of software features regarding data quality issue detection, data quality assessment, and event log reparation. It is
advisable to monitor new publications, e.g., as most of the leveraged R libraries were launched recently. Finally, further research in the field of data quality awareness shall be conducted to point out the importance of event log data quality for the overall success of process mining.
The paper tries to provide a first data-informed taxonomy of Design Thinking. In addition to that it presents a set of prototypes of tools that could be used during data-oriented Design Thinking Workshops, or Data Thinking Workshops (DTW). New products and services increasingly follow a data-driven strategy. This creates the need for designers to create products and services with data in mind at the beginning of the
innovation circle. A common technique used is Design Thinking that has a very narrow view on data so far. Hence, the first part of the paper analyzes the core principles of design processes and derives a data-informed Design Thinking taxonomy. Additionally, the paper provides a set of tools that incorporate this data-informed Design Thinking taxonomy. Using an action design research approach, the DTW format is tested and the results are analyzed using a triangulation approach.
The suggested data-driven Design Thinking taxonomy and the proposed format of DTWs and the provided tools are of potential benefit for designers focusing on developing more digital, data-driven products and services. These help designers sharpen their perspective on data
challenges and come up with a more holistic view on data in their products and services. While the tools suggested in this paper in general work, several newly raised questions and considerations for further research can be derived by the findings. It is hoped this paper will help designers to further develop the taxonomy as well as come up with new tools that can place data at the core of the design process.
The goal of this research is to investigate the use of deep convolutional neural networks for racing bib number recognition in sport images. Several deep neural network architectures are studied. Three final architectures are trained on three different sets of data: 1) Street View House Numbers (SVHN) Dataset, 2) A private dataset from Flashframe.io from different running events, and 3) A combination of dataset 1 and 2. This thesis investigates the performance that can be obtained on racing bib numbers from a neural network that has been trained on solely images from street house numbers, on a mixture of SVHN and RBN images as well as only on RBN images. The motivation behind this is to see how well this problem can be solved by transfer learning, as labelled images of racing bib numbers are scarce.
The models are tested on the RBNR Dataset (Ben-Ami et al., 2012) and a subset of the private dataset from Flashframe.io. The study shows that the best recognition results were obtained by a model trained on the hybrid dataset of all SVHN images plus an additional 50.000 images. This model outperformed the models that had been trained solely on the SVHN Dataset or the private racing bib number dataset.
The best model resulted in Recall of 0,92, Precision of 0,93 and F-measure of 0,93 on the RBNR Dataset (using the same formulas as previously reported on the RBNR dataset), and 0,97, 0,97 and 0,97 on the private dataset, respectively. The reported recognition results on the RBNR dataset are much higher than previously used methods and proves that neural networks can effectively be used for racing bib number recognition.
Deep Learning, a topic of broad and current interest, has undergone rapid development in the last decade. The performance of the algorithms is already above human level. However, the area of Natural Language Processing is still a great challenge for the researchers. This study aims at exploring the novel ULMFiT method for text classification by applying it to a dataset of scientific articles with advanced rhetorical categories. This classification task is challenging even for humans and it requires a substantial analysis of the texts when conducted by machine learning algorithms. The objective of this attempt is to achieve text classification with minimal preparation. The ULMFiT method is the first successful effort to apply transfer learning to NLP tasks. Its performance will be evaluated on a task that requires a level of understanding beyond the semantic meaning of the text.
Companies, public services and other institutions are increasingly turning to web-based applications, but attacks are increasing in both number and variance.
Previous approaches to avoid attacks by using web application firewalls rely primarily on pattern-based detection. This document evaluates if and which machine learning methods can be used to reliably detect web-based attacks. Classifiers such as Support Vector Machines, Neural Networks, Naïve Bayes, Decision Trees and Logistic Regression are used.
Furthermore possible use cases and visualizations of the decisions are suggested.
As an innovative tool, gamification has gained considerable importance in recent years to navigate and find many creative solutions to the problems and challenges faced by existing companies.
In this study, a game toolkit designed by using gamification elements that can potentially serve as a guide to creative ideation is introduced and applied in a data-driven industry. The aim is to
analyze the effects of the specifically designed game toolkit and the applied game elements on creative ideation and the attributes of users in general. The toolkit is explored using insights obtained from the analyses and observation of the behaviors of 16 players during the actual experiments. Furthermore, interviews and online surveys are used to conduct the quantitative and qualitative analyses in order to enhance the game within the scope of Action Design Research.
The toolkit proposed by the researcher is assumed to contribute to the existing literature. It can provide guidance and give valuable suggestions to possible future studies about the creation and
execution of a game toolkit in a field focused on data-driven innovations. Moreover, the results specifically indicate that the game toolkit directly drives creative idea generation while promoting engagement, motivation, and enjoyment as well as altering mindsets and thinking patterns.
Process models are often used as a Knowledge Management method, because these are able to store, visualize, and distribute knowledge within an organization. These knowledge-intensive processes can have a flexible, unstructured form, which is often hard to represent within a process model. The already available approaches of process-oriented Knowledge Management methods, which are considered for Business Process Management, are conducted manually and can therefore become time-consuming and labor-intensive. In addition, in times of Big Data, it is even more challenging to consider all possible cases. Due to the possibility of automatically generating process models through Process Mining, there is a huge potential for creating these processes with real information from event data. This study identifies the possibilities of generating knowledge-intensive process models through Process Mining. At the same time, it addresses the issue of the lack of representation of discovered flexible process models. Although few articles already appeal to the challenge of generating knowledge-intensive processes through Process Mining, the current main focus lies within the operational support in the information systems. The research question of this study is whether actual Process Mining approaches are able to generate process models that are as informative as the models created by processoriented Knowledge Management methods in their current state. To answer the research question, a comprehensive literature review as well as expert interviews have been conducted. Both approaches are part of the design science research methodology, which has been followed during the whole study. The theoretical results have been compared to the practical insights gained from the interviews. As a result, an informal, textual, best practice specification for mining knowledge-intensive process models within a Process Mining project has been developed. This should be considered for implementing an Enterprise Knowledge Medium into the already available IT infrastructure, so that process knowledge can be effectively saved and leveraged through Process Mining.