supervision

Counterfactual explanations (CEs) play a key role in explainable artificial intelligence (XAI), particularly in detecting bias and improving the interpretability of data-driven classification models. A counterfactual explanation identifies the minimal perturbation needed to change the model's decision, thereby shedding light on the reasoning behind the model's decision-making process. However, when applied to statistical learning models, existing counterfactual explanation methods often lack a clear causal structure, leading to explanations that rely on statistical correlations rather than causal reasoning. This makes it difficult to derive explicit and meaningful justifications for the decisions. This study conducts counterfactual reasoning based on a Gaussian Bayesian classifier (GBC), integrating a robust optimization iterative method to identify counterfactual explanations that remain valid even after slight perturbations to the features, thereby improving the reliability and persuasiveness of the counterfactual explanations. Experimental results demonstrate that our method shows strong potential for applications in domains such as medical diagnosis, financial risk assessment, recommendation systems, fairness audits, and legal decision support, providing more causally sound explanations for critical decision-making.

This thesis examines the unintended consequences of the 2008 book ban imposed by the South Korean government on 23 book titles. While book bans are typically intended to suppress the circulation of certain books, they can also increase public attention to the banned books, hence resulting in a reactance phenomenon. Using a panel dataset of monthly book sales from a major South Korean bookstore, this study applies a Difference-in-Differences (DiD) framework using a Two-Way Fixed Effects (TWFE) model to estimate the causal impact of the ban on the sales performance of the banned books compared to non-banned books before and after the ban. The analysis also explores the spillover effect on genre, as well as heterogeneity effects of different buyer characteristics, such as conscription eligibility, gender, political regionalism, and regional GDP. The results suggest that, rather than suppressing demand, the book ban in fact significantly stimulated sales for the banned books. These findings contribute to the literature on consumer behavior and policymaking.

This thesis investigates how the temporary removal of algorithmic amplification, specifically the suspension of trending topics features, impacts citizens’ engagement with political news on major Korean portal platforms during the April 2020 legislative elections. Employing a quasi-natural experimental design and leveraging user-level panel interaction data from over 2,000 users across the two major platforms, the study applies a difference-in-differences framework, supplemented with triple differences, event studies, and machine learning–assisted robustness checks. Results indicate that suspending trending topics significantly reduced user engagement with political articles, particularly among heavy users, without corresponding increases in content diversity or declines in engagement volatility. These findings offer empirical support for the proposition that algorithmic infrastructure influences not just content visibility, but also political attention allocation. Theoretically, the thesis contributes to agenda-setting theory in digital environments and emphasizes the political consequences of de-amplification. Normatively, it raises design questions about democratic information flows, especially in contexts of state–platform interaction. Robustness checks suggest that the effects are not artifacts of selection bias or time-varying confounders.

This study investigates the factors influencing technology acceptance, focusing on the roles of perceived trust, mutual benefits, social influence, and perceived risk. Employing a combination of main effect, mediation, and moderation models, the research aims to elucidate the complex interactions between these variables. The main effect model highlights the critical importance of perceived trust in driving technology acceptance, while the mediation analysis reveals that perceived trust mediates the effects of mutual benefits and social influence on technology acceptance. The moderation analysis examines the influence of task complexity and task importance on the relationship between perceived risk and trust, finding significant moderation by task complexity but not by task importance. The study’s findings underscore the need for managers to prioritize trust-building measures and consider task-related factors when designing strategies to enhance technology adoption. Limitations include the sample size, diversity, and the cross- sectional nature of the data, suggesting the need for future research with larger, more diverse samples and longitudinal designs. Additionally, the current study’s reliance on email advertisements as the treatment medium suggests that future research should explore alternative communication channels to validate the findings. These insights provide a comprehensive understanding of the dynamics affecting technology acceptance and offer practical implications for enhancing adoption in various contexts.

This thesis examines the impact of Paid Family and Medical Leave (PFML) on firm innovation within the United States, employing a novel empirical approach to address a gap in current research. As the U.S. marks the 30th anniversary of the Family and Medical Leave Act (FMLA), which offers unpaid leave, the absence of a federal paid leave policy stands in stark contrast to other OECD countries. This study explores how state-level PFML policies, which provide paid leave, affect firm innovation, a crucial aspect of economic competitiveness that has been underexplored in the context of PFML. Utilizing a difference-in-differences methodology as proposed by Callaway and Sant’Anna, the research investigates firm-level innovation outcomes, specifically the number of patents and adjusted total citations, across various states with staggered implementations of PFML policies from 1994 to 2015. The findings indicate that PFML has a positive and statistically significant effect on firm innovation, suggesting that PFML policies do not merely alleviate employee challenges but actively contribute to enhancing a firm’s innovative output. By integrating PFML into the broader discussion of workforce policies, this study challenges the prevailing view of such policies as burdensome costs to businesses. Instead, it highlights their potential as investments in human capital that yield returns in the form of higher innovation. These insights are particularly valuable for policymakers, business leaders, and scholars as they consider the implications of expanding family-friendly policies to foster a more inclusive and productive economic environment.

This study investigates the impact of Corporate Social Responsibility (CSR) Visibility on financial performance using a quasi-experimental design. CSR Visibility is operationalized through a leading sustainability index called the Dow Jones Sustainability World Index (DJSI). The study focuses on firms that are accepted or declined by a narrow margin of CSR score, which allows for the examination of the effect of DJSI membership on performance through the use of a Regression Discontinuity Design (RDD). Based on my empirical findings, it appears that greater CSR visibility does not result in improved financial performance. This suggests that firms with higher levels of CSR visibility may need to allocate more resources towards CSR activities in order to meet stakeholder expectations, or that stakeholders may not attach significant value to CSR certification. As a result, these firms may experience a decline in financial performance in the short term.

Private Labels (PL) continue to grow in sales and portfolio, driven by the increasing interest from both retailers and consumers toward PL. There has been continuous research in the field, and even after the constant improvement in the quality of PL products, the main concern faced by consumers when deciding whether to switch from National Brands to PL is the perceived risk. We test three different promotions, Price Promotion, Extra Warranty, and Free Returns, to evaluate the effectiveness to increase consumer PL purchase decisions and reduce perceived risk. This paper conducted three studies, to measure the main effect of the promotions on PL purchase decisions, PL purchase intention, and perceived risk. The effect was measured using an online base purchase scenario. A total of 300 participants were collected using Prolific Academic to collect the data. All three strategies were found to positively influence consumer purchase decisions of treated Private labels, varying across product price ranges. Study 3 shows that perceived risk mediates the effect of the promotion strategies. Also proving that results are still significant for “self” and “others” purchase scenarios.