Machine learning and causality: Building robust and scalable models for decision-making

Machine Learning (ML) has witnessed widespread adoption across diverse fields, revolutionizingdecision-making processes. However, ML models, while powerful, often lack robustness, especially when facing data biases and distribution shifts. Causal methods, focusing on understanding and exploiting causal relationships, offer a promising avenue for building robust models. However, the need for experimental or counterfactual data further complicates causal conclusions. Moreover, causal methods are found to scale badly with data dimensions and complexity. This research delves into the intersection of ML and causal inference, aiming to bridge the gap between robustness and scalability. We explore how biases present in training data can lead to suboptimal or even catastrophic decisions, emphasizing the need for models that learn causal relationships without relying on these biases. Further, we strive to also enhance the efficiency and accessibility of causal analysis using ML techniques. The first section of this research addresses biases present in training datasets. We initially focus on solutions to modify training data using causal techniques. We systematically analyze the impact of biases on deep learning models and extend a causal pre-training debiasing technique, emphasizing the importance of accounting for underlying data-generating mechanisms. Additionally, we investigate the benefits of active acquisition methods, exploring their interaction with biases and their role in training robust ML models. In the realm of ML for causality, we tackle the challenge of inferring causal relationships in complex, continuous-time dynamical systems. Traditional interventionist approaches face limitations in dynamic settings, necessitating a data-driven exploration of causation. Our approach focuses on spatiotemporal, system-level thinking, enabling the direct examination of underlying dynamics. We propose a learning paradigm that uses machine learning methods to answer the causal question of why specific events occur in stochastic processes, offering formal and computational tools to uncover and quantify causal relationship

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.