Most machine learning models operate under a comforting but often false assumption: the rules of the game never change. In the world of causal inference, this is known as the stationarity assumption. It posits that if variable $A$ causes variable $B$ today, it will do so in the exact same way tomorrow.
However, real-world systems are rarely so polite. In financial markets, a policy shift can fundamentally alter how interest rates influence inflation. In biology, a maturing organism undergoes structural reorganizations that change how proteins interact. In cloud computing, as noted by Philipp Geiger and Bernhard Schölkopf in Causal models for debugging and control in cloud computing, software updates and hardware degradation constantly violate stationarity, rendering static causal models obsolete.
The research direction emerging from these challenges focuses on Dynamic Causal Discovery: the ability to identify, track, and adapt to causal relationships that evolve over time.
The Failure of Static Graphs
Traditional causal discovery algorithms, such as those reviewed by Simi Job and Xiaohui Tao in Exploring Causal Learning through Graph Neural Networks: An In-depth Review, typically produce a single, fixed Directed Acyclic Graph (DAG). While powerful, these static representations suffer from two primary issues in dynamic environments:
- Causal Drift: This is a specific form of concept drift where the underlying data-generating process evolves. Jeroen Berrevoets and Wouter Verbeke demonstrate in Causal Simulations for Uplift Modeling that when the relationship between a treatment and an outcome shifts, models trained on historical data provide misleading predictions of incremental impact.
- Structural Regimes: Many systems don't just drift smoothly; they jump between different "modes" of behavior. Naoki Chihara and colleagues argue in Modeling Time-evolving Causality over Data Streams that causality in streams like pandemic tracking or stock markets fluctuates between stable periods (regimes), requiring models that can detect these structural shifts in real-time.
From Offline Batches to Online Adaptation
To solve this, researchers are moving causal discovery into the online setting. Instead of treating discovery as a one-time preprocessing step, it becomes a continuous process of hypothesis generation and refinement.
In the context of multi-channel advertising, Alessandro Nuara and Nicola Gatti developed IDIL (IDIL: Exploiting Interdependence to Optimize Multi-Channel Advertising Campaigns). Unlike traditional methods that use a fixed training period, IDIL continuously updates its causality graph to account for shifting dependencies between ad channels, allowing for real-time revenue optimization.
A similar logic applies to decision-making agents. Sanghack Lee and Elias Bareinboim introduced Structural Causal Bandits (Structural Causal Bandits: Where to Intervene?), noting that the stationarity assumption fails when the mechanisms governing actions and rewards evolve. By identifying the "possibly-optimal minimal intervention set" (POMIS), agents can prune redundant actions even as the environment changes.
Identifying Invariant Mechanisms
One of the most intuitive ways to handle non-stationarity is to look for what doesn't change. Khurram Javed, Martha White, and Yoshua Bengio propose a method in Learning Causal Models Online that identifies causal structures by removing "spurious" features—those whose correlation with a target fluctuates over time.
The intuition is that while correlations might shift due to external factors, the true causal mechanism should remain relatively stable. By tracking the variance of feature weights in an online stream, their algorithm can achieve zero-shot generalization across shifting distributions without needing explicit environment labels.
Context-Aware Discovery
Sometimes, the "change" in causality is actually a response to an internal state. Wiebke Günther and Martin Rabel address this in Causal discovery with endogenous context variables through an algorithm called PC-AC (Adaptive Context).
Instead of assuming one graph fits all data, PC-AC adaptively switches between:
- Pooled data: Used when the causal mechanism is invariant across contexts.
- Context-specific data: Used when the existence or strength of a relationship depends on the current regime.
This prevents the "selection bias" that occurs when researchers try to force a single, static DAG onto a system that behaves differently in different states.
Applications in Complex Systems
The shift toward dynamic causality is particularly visible in fields where "static" is not an option:
- Climate Science: Dongqi Fu and colleagues developed TacSas (Generating Fine-Grained Causality in Climate Time Series Data for Forecasting and Anomaly Detection) to model climate patterns where interactions are non-linear and evolve rapidly, outperforming static methods in predicting rare anomalies like thunderstorms.
- Manufacturing: Johannes Huegle and Rainer Schlosser argue in Understanding unforeseen production downtimes in manufacturing processes using log data-driven causal reasoning that because machinery degrades over time, causal models must be updated online to accurately identify the root causes of downtime.
- Fairness in Streams: Leizhen Zhang and Yi He introduced SFCF (Fairness-Aware Streaming Feature Selection with Causal Graphs), which integrates drift detection to ensure that as causal structures change, the model doesn't inadvertently begin using new "proxy" features that leak sensitive information.
By treating the causal graph as a living, breathing structure rather than a static artifact, this research direction enables AI systems to maintain their reasoning capabilities in a world that is constantly in flux.