Refine
Document Type
- Master's Thesis (10)
- Article (7)
- Bachelor Thesis (7)
- Book (1)
- Lecture (1)
Language
- English (26) (remove)
Analysing the systematics of search engine autocompletion functions by means of data mining methods
(2017)
In the internet era, the information that can be found about politicians online can influence
events such as the results of elections. Research has shown that biased search rankings can
shift the voting preferences of undecided voters. This shows the importance of studying
online search behaviour, especially in the pre-elections phase, when search results can
have a particular influence on the future political scene of a country.
This master thesis aimed to study the behaviour of online search engines in a period before
the German federal election in 2017. The aim was to ascertain if there is any pattern to be
found in the auto-suggestions for searches related to politicians.
In order to gather data for this experiment, a crawler browsed search engine web pages,
input a name and a surname of a politician, and saved that together with all autosuggestions
from the search engine. The autosuggestions were prepared for the analysis and
divided into semantic groups with the help of clustering algorithms.
Different statistical methods, such as correlation analysis, regression analysis, and clustering
were used to identify patterns in the data. The research showed that there are
no particularly strong patterns in the autosuggestions for searches related to politician’s
names. Only moderate dependence was found between gender and personal topics, and
showed that a higher amount of personal information autosuggestions correspond more
to female politicians.
The goal of this work is to detect "gender biases" in the communication of users of Subreddits on the platform Reddit. The analysis is carried out for eleven selected Subreddits. Furthermore, an attempt is made to identify different user types with the help of a k-means clustering and also to analyze "gender biases" in their communication. Based on the aggregated datasets, fasttext Word Embedding models are trained to identify terms that show high semantic relatedness in terms of cosine similarity of their word vectors with selected feminine and masculine terms.
To this end, the terms are analyzed for sentiment using the NRC-VAD Lexicon and tested for statistically significant differences. In addition, the Word Embedding Association Test (WEAT) is performed to check for subliminal associations. In relation to the considered text corpus, it is essentially observed that women are frequently associated with adjectives that associate them with appearances,
childbearing abilities or adaptability also in relation to the family. In contrast, men are associated with and measured by adjectives that refer to their prestige, strengths and weaknesses, career or physical characteristics.
As the information era progresses, the sheer volume of information calls for sophisticated retrieval systems. Evaluating them holds the key to ensuring the reliability and relevance of retrieved information. If evaluated with renowned methods, the measured quality is generally presumed to be dependable. That said, it is often forgotten that most evaluations are only snapshots in time and the reliability might be only valid for a short moment. Further, each evaluation method makes assumptions about the circumstances of a search and thereby has different characteristics. Achieving reliable evaluation is critical to retain the aspired quality of an IR system and maintain the confidence of the users. Therefore, we investigate how the evaluation environment (EE) evolves over time and how this might affect the effectiveness of retrieval systems. Further, attention is paid to the differences in the evaluation methods and how they work together in a continuous evaluation framework. A literature review was conducted to investigate changing components which are then modeled in an extended EE. Exemplarily, the effect of document and qrel updates on the effectiveness of IR systems is investigated through reproducibility experiments in the LongEval shared task. As a result, 11 changing components together with initial measures to quantify how they change are identifed, the temporal consistency of five IR systems could precisely be quantifed through reproducibility and replicability measures and the findings were integrated into a continuous evaluation framework. Ultimately, this work contributes to more holistic evaluations in IR.
The purpose of this research lies in uncovering the participants emotions when watching a personalized advertisement on the social media Instagram. This is of use to the marketing and psychology research community to discover more on consumer behavior and the controversy between privacy concerns and usefulness of advertisement personalization. The research question reads: “Does the use of personalization on social media advertisements incite (1) a change in the emotional state and (2) recall capability of German Instagram users aged 18-30 that diverges from the psychophysiological parameters measured by exposing these users to the same advertisements without personalization?”
Psychophysiological tests are used in combination with two self reported questionnaires that assess the participants positive and negative effect and the recall and recognition differences between the group given personalized stimuli including the participants name, location, and activity and the one group given impersonalized ones. The sample consists of n=31 German-speaking participants between the age of 18 and 30.
The results, although not all of statistic relevancy of α=0,05, show a trend that personalized advertisements instigate more positive valence and activation as not personalized stimuli. No significant or trending difference was found to the recall and recognition capabilities of the two groups.
Kein Abstract vorhanden.
Digital curation is currently not very well covered by university curricula in the German speaking countries. Nevertheless there is a strong demand for well-educated staff in this field. As part of the project “nestor”, a transnational partnership of academic institutions in Germany, Switzerland, and Austria, a comprehensive qualification program based on e-learning tutorials, schools, seminars, and publications has been established to meet this demand.
This paper examines different business models of companies dealing with (earmarked) remittances and sheds light on the associated challenges of the industry, specifically, remittances for health, based on the model of the fintech startup GloryHealthCare. The work "Business Model Generation" by Osterwalder and Pigneur (2010) is used as a method for the analysis, as this is often used as a basis for the business models of startups. The study focuses regionally on Europe and Africa, as Germany and Ghana are the start-up's first target markets. Among other things, the industry's processes, pricing, and existing competition are examined. The SWOT analysis methodology clarifies the individual companies' opportunities and risks and makes a competitive position visible. Meanwhile, network effects of the diverse business models are made visible based on the paper "Digital Economy and Network Effects" by Frank Linde (2020). Network effects play a crucial role in the reach, influence, and competitiveness of existing and new businesses in the remittance industry. The study also emphasizes the importance of knowledge and networks, which are more important than financial resources. The previous aspects considered a basis for developing a new concept as an alternative to the Business Model Canvas: the iBusiness Model. The results of this study provide insights into the design of efficient business models and support companies in the remittance industry in developing strategies to overcome challenges and take advantage of opportunities.
As a result of the widespread use of online networking sites, the ways in which people connect and network, both personally and professionally, have been transformed in recent years. Platforms such as LinkedIn or XING have profoundly changed the dynamics of professional networking by providing new means of contact and creating an environment that promotes the exchange of knowledge and ideas. However, compared to social network sites, professional network sites have received little attention in research despite their growing importance. Particularly, the relationship between the use of professional network sites and users' well-being has been understudied. However, the investigation of these platforms is of societal relevance given their consistent growth and the increasing importance of these platforms for both individuals and companies. Existing research on the relationship between the use of social network sites (SNS) and the subjective well-being of users has identified the usage type (active and passive use) as a relevant variable. The aim of this study was to transfer these findings to the context of professional network sites and to explore the relationship between the type of use of a professional network site and the subjective well-being of its users.For this purpose, the active-passive model of SNS use was applied to the context of professional network sites for the first time. To answer the research question, a quantitative online survey was conducted with 526 LinkedIn users. Results of the mediation analyses revealed an indirect positive relation between active use of LinkedIn and well-being. Conversely, a negative indirect relation was found between passive use of LinkedIn and subjective well-being. All tested mediating variables, including social capital for active use and upward social comparison, downward social comparison and envy for passive use, were determined to be relevant in explaining the link between well-being and active and passive LinkedIn use, respectively.
This thesis aims to extend an existing Open Educational Resource (OER), which is available as a GitHub repository, and provide an organized introduction to basic machine learning (ML) concepts and algorithms. Further models, followed by structured metadata for each object, will be included while adhering to the contribution guidelines of the OER and following the CC license. The Machine-Learning-OER Basics repository intends to provide a wide range of benefits by enabling diverse users to apply and distribute machine learning algorithms. The goal of this digital collection is to fill the existing gap for instructional material on using machine learning in OER as well as make it easier to learn ML concepts effectively. These ML models are developed using the programming language Python and the library scikit-learn, among other standard libraries. Jupyter Notebook will make it straightforward for the user to explore the code. In order to apply the models to various practical scenarios, a non-specific data set is selected. This work is considered a solution approach in that it includes adding classification models.
A performance comparison of the models is conducted. This comparative analysis evaluates the efficiency of each model. The examination includes various metrics for measurement. This work serves as a written extension, providing comprehensive background information on the algorithms utilized within the repositories and the performance comparison.