A Brief Introduction to Temporality and Causality

Causality is a non-obvious concept that is often considered to be related to temporality. In this paper we present a number of past and present approaches to the definition of temporality and causality from philosophical, physical, and computational points of view. We note that time is an important ingredient in many relationships and phenomena. The topic is then divided into the two main areas of temporal discovery, which is concerned with finding relations that are stretched over time, and causal discovery, where a claim is made as to the causal influence of certain events on others. We present a number of computational tools used for attempting to automatically discover temporal and causal relations in data.

💡 Research Summary

The paper “A Brief Introduction to Temporality and Causality” offers a panoramic view of two concepts that are often conflated but have distinct theoretical underpinnings and practical implications. It begins by tracing the philosophical origins of temporality—from Aristotle’s notion of time as a measure of change to Kant’s transcendental intuition—followed by a concise overview of physical theories: Newtonian absolute time, Einsteinian relativistic spacetime, and the more nuanced, observer‑dependent treatment of time in quantum mechanics. These perspectives lay the groundwork for understanding why time appears as a fundamental axis in data, whether in regular time‑series, event logs, or irregular streams such as social‑media feeds.

The discussion then shifts to causality. After a brief historical sketch of causal reasoning in classical philosophy (Aristotle’s four causes, Hume’s regularity view), the authors focus on modern formalizations: Bayesian networks, structural causal models (SCM), and the counterfactual framework popularized by Pearl. Two pillars—intervention (do‑operator) and counterfactual reasoning—are highlighted as the criteria that separate genuine causal claims from mere correlations.

A central contribution of the paper is the clear separation of “temporal discovery” and “causal discovery.” Temporal discovery concerns the identification of patterns that unfold over time. Classical techniques such as cross‑correlation, Granger causality tests, and Dynamic Time Warping (DTW) are described, followed by recent deep‑learning approaches like Temporal Convolutional Networks (TCN), Transformers, and the Temporal Fusion Transformer (TFT). Causal discovery, by contrast, aims to infer directed influence among variables. The authors review constraint‑based algorithms (PC, FCI), score‑based methods (GES), and hybrid approaches, as well as effect‑estimation tools such as instrumental variables, randomized controlled trials, and software libraries like DoWhy and CausalNex.

The paper proceeds to catalog the most widely used computational tools. TETRAD provides a graphical interface for building and testing causal graphs; DoWhy offers a Pythonic pipeline that integrates identification, estimation, and refutation; CausalNex focuses on Bayesian network learning with a strong emphasis on interpretability. For temporal analysis, Prophet (by Facebook) and Kats (by Amazon) deliver automated forecasting, while libraries such as tslearn and the PyTorch‑based TFT implementation enable more flexible, deep‑learning‑driven temporal pattern extraction. The authors compare each tool’s assumptions, data‑type requirements, scalability, and typical application domains.

A forward‑looking section introduces the emerging paradigm of “temporal‑causal hybrid models.” The authors argue that many real‑world systems—clinical decision support, financial risk management, climate dynamics—cannot be adequately described by treating time and causality in isolation. They outline several research directions: (1) integrating structural causal graphs with time‑varying coefficients; (2) employing reinforcement learning to actively probe causal mechanisms while respecting temporal constraints; (3) developing Bayesian non‑parametric time‑series models that capture both stochastic dynamics and causal interventions.

In conclusion, the paper emphasizes that a unified treatment of temporality and causality can dramatically improve the reliability of data‑driven insights. It calls for future work on scalable algorithms for high‑dimensional data, rigorous quantification of causal uncertainty, and ethical frameworks for interpreting causal claims—especially when automated systems influence high‑stakes decisions. By bridging philosophical foundations, physical theory, and state‑of‑the‑art computational methods, the article serves as a concise yet comprehensive roadmap for researchers and practitioners seeking to uncover “when” and “why” in complex datasets.