Contributions in Merlin

This master thesis explores domain adaptation methods for pre-trained Large Language Models (LLMs) to map natural language mentions from a text genre onto a target domain ontology based on cosine similarity in a semantic vector space. For the thesis, the input mentions are skill requirement mentions extracted from Swiss job ad postings written in German or English, and the target domain onto which these terms have to be mapped is the European Skills, Competences, Qualifications and Occupations (ESCO) ontology. The objective of this task is to track changes in the labor market and help recruiters fill positions based on skill requirements fulfilled by candidates. The thesis explores three methods: Masked Language Modelling, Multiple Negative Ranking Loss, and binary classification method for further pre-training in order to adapt LLMs to a target domain ontology. Experiments were conducted on 15 model variants using different input data and starting models. Two gold standard datasets, one consisting of randomly selected skill requirement mentions, and the other specifically crafted from challenging cases, were used for evaluating model performance. The evaluations were created by annotating the top suggestions made by our model variants. Mean Average Precision (MAP) scores were computed based on human annotations of the suggestions, made by each model variant for each term in the gold standard datasets. MAP is used as an evaluation metric since more than one mapping might be correct or acceptable, and a good ranking of the appropriate ontology concepts can be measured via this metric. The MNR models with the hard negative sampling strategy, wherein the negative samples are taken with lexical and semantic similarities to the anchor term, and domain adaptation on both the job-ads data and the ESCO ontology data were found to be the best-performing model variants for both the English and German languages. The thesis concludes that domain adaptation on both the input texts and the target domain is beneficial for mapping mentions from the input genre onto the target domain. It also suggests that using a hard negative sampling method for creating the MNR data is beneficial compared to a random negative sampling method.

This work aims to investigate the feasibility of exploiting reinforcement learning (RL) to improve the impact of ransomware on a target device while evading dynamic detection methods such as behavioral fingerprinting-based anomaly detection (AD). Given the constantly growing number of connected resource-constrained devices, such as Internet of Things (IoT) devices, and the significant rise in ransomware attacks over the past years, the importance of investigating ransomware attacks and corresponding defense approaches is evident. So far, most related research has been confined to exploring unethical artificial intelligence (AI) systems instead of analyzing the possibilities of using AI for launching optimized malware attacks. This work covers the mentioned limitations and introduces Ransomware Optimized with AI for Resource-constrained devices (ROAR), an RL framework to hide ransomware from dynamic detection mechanisms and optimize its impact on the target device. ROAR has been deployed in a real-world IoT crowdsensing scenario, including a Raspberry Pi 4 as a spectrum sensor. The Raspberry Pi was infected with ROAR, and behavioral data were collected from the target device to facilitate environment simulation. The results obtained by executing prototypes of the RL agent have been aggregated, and the corresponding plots are discussed and compared. These findings suggest that no relation exists between individual actions within an episode and that discounting future rewards does not improve performance in this particular RL problem. Overall, this work demonstrates the feasibility of optimizing ransomware attacks with RL and the effectiveness of the resulting evasion capabilities. The findings derived from the collected results hold in a simulated environment and when the agent is deployed in a real scenario. To our knowledge, this work is the first to investigate the possibilities of supporting malware attacks with RL during the attack phase. Further studies are needed to investigate additional optimizations of the RL model, efficiency improvements to the underlying ransomware implementation, and the feasibility of attacking more powerful devices.

We aim to enhance the performance of the standard dollar carry by harnessing news sentiment. Our study focuses on determining the impact of news sentiment on the average spot returns of the G10 currency basket against the USD at the monthly level. The empirical analyses reveal the existence of a predictive relationship between the two variables. Additionally, we observe that the effect of USD news sentiment varies across different quantiles of the average spot returns. Given that, we propose two news sentiment enhanced dollar carry strategies. The first strategy exploits the result of our initial findings, while the second one further leverages information from both of our research insights. The predictability manifests itself in the form of higher Sharpe ratios of our two augmented dollar carry strategies over both the whole sample period (February 2000 to March 2022) and the out-of-sample period (January 2014 to March 2022) than that of the standard dollar carry when transaction costs are disregarded. Notably, during the out-ofsample period, the second strategy generates a Sharpe ratio of 0.48 and the first strategy 0.43, compared to the benchmark’s -0.22. Despite the fact that transaction costs can eat up a large part of the profits, our findings suggest that news sentiment still has the ability to improve the dollar carry strategy by achieving excess returns. This is evident in the out-of-sample period, where news sentiment based strategies continue to outperform the benchmark.

The number of collisions in the upcoming runs of the Large Hadron Collider at CERN will increase significantly. The increasing amount of data and a higher granularity of the newly developed calorimetric detectors pose a substantial data volume and complexity challenge to the current particle shower reconstruction algorithms. This thesis aims to explore the feasibility of machine-learned models scalable to large data volumes for improving the reconstruction quality of calorimetric particle showers via calorimetric clustering. The goal of calorimetric clustering is to recognize and reconnect fragmented energetic components of particle showers described by three-dimensional spatial structures called tracksters. We show that machine-learned models are viable methods for calorimetric clustering and provide a significant reconstruction performance benefit over classical clustering approaches. Furthermore, we investigate the feasibility of node classification and link prediction problem formulations for training graph neural networks. Experimentally, we show that graph-based models provide a better reconstruction performance, more compact data representation, and better scalability on the tested datasets than feed-forward neural networks.

null

Although influencer marketing has been experiencing tremendous growth, and many firms have been trying to find the right influencers for their marketing campaigns, there has been little research on identifying the timing of the high-impact work or analyzing the hot streak periods of influencers on social media. My thesis is among the first works in this area on the TikTok platform where I applied some methods used in previous literature. All the analyses were made using the video-level popularity metrics of TikTok: “number of likes”, “number of shares”, “number of plays”, and “number of comments”. In addition, unlike other studies, I created the visualization of the hot streak durations of TikTok authors to analyze information such as the start, and end time of hot streaks, or the length of hot streaks for each author. All plots for the analyses were built with bokeh- a Python library for interactive data visualizations. In conclusion, with the dataset in my thesis study, first I found that the timing of the most popular video is random in TikTok users’ lifecycles. Second, I discovered that the timing of the biggest hit and the second hit are close to each other, so TikTok authors may experience average success close to their most popular videos. Third, I detected a pattern indicating that TikTok authors use more diverse hashtags during hot streak periods than before them. Lastly, I observed that the relative hot streak length -success duration- of an author is usually between 12% and 50%. As a result, marketers can use these observations in my thesis to identify successful influencers on the TikTok platform. Finally, the formulas and visualization methods used in my thesis can be applied to a larger TikTok dataset or other datasets from other social media platforms such as Instagram, or YouTube.

During the global financial crisis, a detachment of the LIBOR from banks’ borrowing costs led to investigations, revealing severe misreporting of quotes in the fixing methodology. Subsequently, the LIBOR was discontinued and replaced by the SARON in Switzerland. By estimating bank level average borrowing costs of major Swiss banks and applying a Difference-in-Differences model, this thesis analyzes their evolution and drivers. Monetary policy, market risk and the term structure of liabilities significantly influence borrowing costs, while no advantage for too-big-to-fail banks is identified. In addition, SARON is more correlated to the estimated borrowing costs than LIBOR, indicating a better fit.

null

Fintech and big tech companies are making rapid inroads into credit markets. We hand construct a global database of fintech and big tech lending volumes for 79 countries over 2013-2018. Using a panel regression analysis, we find these new forms of digital lending are larger in countries with higher GDP per capita (albeit at a declining rate), where banking sector mark-ups are higher, and where banking regulation is less stringent. We also find that these alternative forms of credit are more developed where the ease of doing business is greater, investor protection disclosure and the efficiency of the judicial system are more advanced, and where bond and equity markets are more developed. Overall, fintech and big tech credit seem to complement other forms of credit, rather than substitute for them.

Affect is involved in many psychological phenomena, but a descriptive structure, long sought, has been elusive. Valence and arousal are fundamental, and a key question–the focus of the present study–is the relationship between them. Valence is sometimes thought to be independent of arousal, but, in some studies (representing too few societies in the world) arousal was found to vary with valence. One common finding is that arousal is lowest at neutral valence and increases with both positive and negative valence: a symmetric V-shaped relationship. In the study reported here of self-reported affect during a remembered moment (N = 8,590), we tested the valence-arousal relationship in 33 societies with 25 different languages. The two most common hypotheses in the literature–independence and a symmetric V-shaped relationship–were not supported. With data of all samples pooled, arousal increased with positive but not negative valence. Valence accounted for between 5% (Finland) and 43% (China Beijing) of the variance in arousal. Although there is evidence for a structural relationship between the two, there is also a large amount of variability in this relation.