NSBE - Business Analytics
URI permanente para esta coleção:
Navegar
Entradas recentes
- Machine learning-based insolvency prediction for Siemens’ extended payment termsPublication . Otter, Felix; Obermeier, DanielThis thesis investigates short-term insolvency prediction using solely behavioral payment data and explainable AI in the context of Siemens' Extended Payment Terms program. A dual modeling approach combines interpretable baselines (logistic regression, random forest, and XGBoost) with a sequence-based Long Short-Term Memory (LSTM) model. The LSTM achieves moderate discrimination at natural default prevalence, while SHAP-based explanations and interviews with risk managers show that cross-model consistency and local explanations strengthen trust in the model’s outputs.
- Large language models (LLMs) for legal analysis: RAG and beyond for optimizing domain adaptation in Portuguese legal domainPublication . Barros, Tiago Mendonça Alencar; Han, QiweiThis study explores RAG systems tailored to the Portuguese legal domain, highlighting challenges in underrepresented languages. Fixed-size chunking strategies, particularly Token Text Splitter, were found to be most effective, while more advanced techniques like Recursive and Semantic splitting showed little benefits. Larger chunk sizes improved retrieval accuracy and answer quality, though the impact of chunk overlap remains inconclusive. Self-reflection techniques show promising results, particularly for weaker LLMs. Techniques such as adding a pre-post translation proved to be an efficient technique for mitigating language bias.
- Large language models (LLMS) for legal analysis rag and beyond for optimizing domain adaptation in Portuguese legal domain-advanced retrieval and prompt engineering techniques for retrieval augmented generation in the Portuguese legal domainPublication . Thaçi, Rita; Han, QiweiThis study explores RAG systems tailored to the Portuguese legal domain, highlighting challenges in underrepresented languages. Fixed-size chunking strategies, particularly Token Text Splitter, were found to be most effective, while more advanced techniques like Recursive and Semantic splitting showed little benefits. Larger chunk sizes improved retrieval accuracy and answer quality, though the impact of chunk overlap remains inconclusive. Self reflection techniques show promising results, particularly for weaker LLMs. GraphRAG shows promise with faster results than traditional RAG approaches. Reranking techniques can improve retrieval but require larger, diverse datasets. Reasoning-based and zero-shot prompting improve accuracy in multi-hop scenarios. Pre-post translation proved to be an efficient technique for mitigating language bias.
- Examining political polarization in the German bundestag using large language models: historical trends and a contemporary analysis - bert modelPublication . Amenda, Anna Charlotte; Shen, YufeiAnalyzing political polarization has become increasingly relevant, particularly in light of the recent government crisis in Germany. This research investigates how political polarization in Germany has evolved over time and identifies factors influencing polarization in the currente lectoral term (2021-2025). We utilize an ensemble of three Large Language Models, BERT, GPT-4o-mini, and LLaMA, to classify speeches in the German Bundestag as polarizing. This approach is complemented by sentiment and structural analysis. Our results show a significant increase in political polarization across the last two electoral terms, with the entry of the right-wing party Alternative für Deutschland (AfD) into the Bundestag occurring concurrently. Political parties, followed by topics discussed, have emerged as the most influential factors in polarization. Meanwhile, the recent dissolution of the governing coalition was only subtly indicated by a reduction of applause among governing parties.
- Examining political polarization in the German bundestag using large language models: historical trends and a contemporary analysis - sentiment and structural analysisPublication . Bienert, Silja Sophie; Shen, YufeiAnalyzing political polarization has become increasingly relevant, particularly in light of the recent government crisis in Germany. This research investigates how political polarization in Germany has evolved over time and identifies factors influencing polarization in the current electoral term (2021-2025). We utilize an ensemble of three Large Language Models, BERT, GPT-4o-mini, and LLaMA, to classify speeches in the German Bundestag as polarizing. This approach is complemented by sentiment and structural analysis. Our results show a significant increase in political polarization across the last two electoral terms, with the entry of the right-wing party Alternative für Deutschland (AfD) into the Bundestag occurring concurrently. Political parties, followed by topics discussed, have emerged as the most influential factors in polarization. Meanwhile, the recent dissolution of the governing coalition was only subtly indicated by a reduction of applause among governing parties
- Examining political polarization in the German bundestag using large language models: historical trends and a contemporary analysis - Gpt modelPublication . Thomas, Antoine Otto; Shen, YufeiAnalyzing political polarization has become increasingly relevant, particularly in light of the recent government crisis in Germany. This research investigates how political polarization in Germany has evolved over time and identifies factors influencing polarization in the current electoral term (2021-2025). We utilize an ensemble of three Large Language Models, BERT, GPT-4o-mini, and LLaMA, to classify speeches in the German Bundestag as polarizing. This approach is complemented by sentiment and structural analysis. Our results show a significant increase in political polarization across the last two electoral terms, with the entry of the right-wing party Alternative für Deutschland (AfD) into the Bundestag occurring concurrently. Political parties, followed by topics discussed, have emerged as the most influential factors in polarization. Meanwhile, the recent dissolution of the governing coalition was only subtly indicated by a reduction of applause among governing parties
- Big data meets the big screen: a comprehensive machine learning study about the movie industry-understanding box-office dynamics: the role of timing and audience in movie revenue predictioPublication . Gonçalves, Tomás Shoemaker Simão; Guha, SreyaaThis thesis investigates factors influencing movie success using machine learning and deep learning techniques. Traditional machine learning methods analyze key determinants of box office performance, such as audience and critic ratings, release timing, and sequel dynamics. Deep learning approaches, including Natural Language Processing and Time Series Classification, examine patterns in sequential movie data. By integrating diverse data sources and predictive modeling techniques across the movie lifecycle, this study provides a comprehensive perspective on audience engagement and market dynamics. The findings advance academic understanding of the topic and offer actionable recommendations for industry stakeholders aiming to optimize performance.
- Adversarial generative forecasting of daily Fraud Amount for sparse transaction time seriesPublication . Mueller, Moritz; Xufre, PatríciaIn collaboration with SIBS, this project forecasts daily accepted fraud amounts in e-commerce transactions to support proactive risk management. Utilizing a dataset ofover 166 million transactions (2023–2024), we engineered behavioral features to benchmark multiple machine learning models. XGBoost was the champion model, achievinga 16.56% MAPE, but struggling during volatility spikes. To improve reliability duringspike periods, we investigated three complementary strategies: GAN-based data augmentation to increase exposure to synthetic high-fraud scenarios, quantile-based forecasting to model the upper tail of the distribution, and transfer learning approaches that adapt large pre-trained time-series models to our use case.
- Using machine learning to Solve Real Banking challenges at Banco PrimusPublication . Rudolf, Nils; Batikas, MichailThis thesis examines how machine-learning models and explainable AI can be used to analyze two distinct use cases: loan conversion and cross-selling in retail banking. Using proprietary data from Banco Primus, logistic regression, random forest, and XGBoost models are evaluated using business-oriented back-testing. SHAP is applied to explain predictions and identify key drivers. Approved loan conversion is mainly associated with partner characteristics, process timing, and communication availability. Personal loan cross-selling is mainly associated with external credit profiles and behavioral history, revealing campaign fatigue. The findings support process optimization in loan origination and propensity-based targeting frameworks for cross-selling.
- Predicting approved loan conversionPublication . Pachelbel, Valeska Von; Batikas, MichailThis thesis examines how machine-learning models and explainable AI can be used to analyze two distinct use cases: loan conversion and cross-selling in retail banking. Using proprietary data from Banco Primus, logistic regression, random forest, and XGBoost models are evaluated using business-oriented back-testing. SHAP is applied to explain predictions and identify key drivers. Approved loan conversion is mainly associated with partner characteristics, process timing, and communication availability. Personal loan cross-selling is mainly associated with external credit profiles and behavioral history, revealing campaign fatigue. The findings support process optimization in loan origination and propensity-based targeting frameworks for cross-selling.