Semantische Suche

In this thesis, we investigated the automatic classification of scientific papers in the Software Engineering domain. In doing so, a classification scheme of paper contents with regard to Research Object, Statement, and Evidence was consolidated. We then investigate in a comparative analysis the machine learning algorithms Naïve Bayes, Support Vector Machine, Multi-Layer Perceptron, Logistic Regression, Decision Tree, and BERT applied to the classification task.

There are multiple examples that highlight the importance of dependency estimation, like knowing there exists a correlation between the regular dose of a drug and the health of a patient helps to understand the impact of a newly manufactured drug. Knowing how the case material, brand, and condition of a watch influences the price on an online marketplace can help to buy watches at a good price. Material sciences can also use dependency estimation to predict many properties of a material before it is synthesized in the lab, so fewer experiments are necessary.

Many dependency estimation algorithms require a large amount of data for a good estimation. But data can be expensive, as an example experiments in material sciences, consume material and take time and energy. As we have the challenge of expensive data collection, algorithms need to be data efficient. But there is a trade-off between the amount of data and the quality of the estimation. With a lack of data comes an uncertainty of the estimation. However, the algorithms do not always quantify this uncertainty. As a result, we do not know if we can rely on the estimation or if we need more data for an accurate estimation.

In this bachelor's thesis we compare different state-of-the-art dependency estimation algorithms using a list of criteria addressing these challenges and more. We partly developed the criteria our self as well as took them from relevant publications. The existing publications formulated many of the criteria only qualitative, part of this thesis is to make these criteria measurable quantitative, where possible, and come up with a systematic approach of comparison for the rest.

From 14 selected criteria, we focus on criteria concerning data efficiency and uncertainty estimation, because they are essential for lowering the cost of dependency estimation, but we will also check other criteria relevant for the application of algorithms. As a result, we will rank the algorithms in the different aspects given by the criteria, and thereby identify potential for improvement of the current algorithms.

We do this in two steps, first we check general criteria in a qualitative analysis. For this we check if the algorithm is capable of guided sampling, if it is an anytime algorithm and if it uses incremental computation to enable early stopping, which all leads to more data efficiency.

We also conduct a quantitative analysis on well-established and representative datasets for the dependency estimation algorithms, that performed well in the qualitative analysis. In these experiments we evaluate more criteria: The robustness, which is necessary for error-prone data, the efficiency which saves time in the computation, the convergence which guarantees we get an accurate estimation with enough data, and consistency which ensures we can rely on an estimation.

Battery based load hiding gained a lot of popularity in recent years as an attempt to guarantee a certain degree of privacy for users in smart grids. Our work evaluates a set of the most common privacy measures for BBLH. For this purpose we define logical natural requirements and score how well each privacy measure complies to each requirement. We achieve this by scoring the response for load profile altering (e.g. noise addition) using measures of displacement. We also investigate the stability of privacy measures toward load profile length and number of bins using specific synthetic data experiments. Results show that certain private measures fail badly to one or many requirements and therefore should be avoided.

However, previous research showed that Load Disaggregation underperforms in the industrial setting compared to the household setting. Also, the domain knowledge available about industrial processes remains unused.

The objective of this thesis was to improve load disaggregation algorithms by incorporating domain knowledge in an industrial setting. First, we identified and formalized several domain knowledge types that exist in the industry. Then, we proposed various ways to incorporate them into the Load Disaggregation algorithms, including Theory-Guided Ensembling, Theory-Guided Postprocessing, and Theory-Guided Architecture. Finally, we implemented and evaluated the proposed methods.