Abstract
This dissertation explores the integration of causal inference frameworks into the study and development of Recommender Systems, with the goal of improving their robustness, interpretability, and scientific rigour. Traditional Recommender Systems often rely on observational data and statistical correlations, which can lead to biased estimates and unreliable outcomes when deployed in real-world environments. To address these challenges, we propose a causal perspective grounded in both the Potential Outcomes and Structural Causal Models frameworks. We demonstrate how this approach enables formal definitions of key estimands, such as total, direct, and indirect effects, and supports the identification and mitigation of confounding and selection biases. Furthermore, the dissertation investigates the issue of reproducibility in the Recommender Systems and Rein forcement Learning literature, highlighting significant shortcomings in current research practices. Our empirical analysis reveals that only a small fraction of published Recommender Systems papers meet basic reproducibility standards, which hinders scientific progress. To support more reliable and transparent experimentation, we present a detailed taxonomy of reproducibility artifacts and propose actionable guidelines for the research community. Finally, we introduce a unified simulation framework for modelling user-system interactions in RSs. This framework integrates point processes for modelling user activity over time, Markov models for navigation behaviour, and Structural Causal Models for modelling user choices. Through a series of illustrative scenarios, we show how even simple, knowledge-driven models can approximate real-world behaviours while preserving interpretability and explicit control over assumptions. By bringing together causality, reproducibility, and simulation, this work aims to bridge gaps in Recommender System research and promote a more principled foundation for understanding and designing recommendations. The thesis not only addresses the theoretical and methodological limitations of current approaches but also provides empirical insights and practical guidelines to support the development of more robust, interpretable, and reliable recommendation algorithms.