Traditional causal discovery operates under a comforting assumption: stationarity. It assumes that the underlying rules governing a system remain constant over time (Samantha Kleinberg et al.). Whether mapping biological processes, financial markets, or physical systems, standard models assume a single, invariant global graph applies across all data points (Inwoo Hwang et al.).
However, real-world systems are rarely static. Relationships evolve as new conditions emerge, external policies shift, or components physically alter their connections mid-sequence (Sriraam Natarajan et al.; Causal Discovery in Physical Systems from Videos). When a system encounters "concept drift"—where the mapping between interventions and outcomes shifts sequentially—static causal models fail to track the moving optimum (Sam Verboven; Virginia Aglietti et al.). To maintain accuracy without resorting to computationally expensive full-model retraining, we need frameworks that treat causal structures as dynamic, evolving entities (Alessandro Nuara et al.).
Turning Heterogeneity into a Signal
In classical statistics, data heterogeneity or distribution shifts are often treated as a nuisance to be cleaned away. However, in non-stationary causal discovery, this "messiness" becomes the primary signal.
As Fangting Zhou et al. argue, when causal relationships change across different environments or conditions, tracking those changes allows us to unlock previously unidentifiable causal structures. By modeling the causal structure directly, systems can adapt to temporal shifts by identifying which variables are the active drivers of change versus those that are merely affected by it (Kun Zhang et al.).
Localizing Causal Graphs via Regimes and Contexts
One powerful way to handle non-stationarity is to break a global system down into local, context-specific regimes. Instead of forcing a single graph onto the entire dataset, models can adaptively switch between pooled and context-specific data to capture fine-grained details without introducing selection bias (Wiebke Günther et al.).
Abdellah Rahmani demonstrates this with the CASTOR framework, which simultaneously identifies hidden operational regimes and learns the specific causal graph governing each one. This is crucial because regime boundaries are fundamentally defined by shifts in the underlying causal mechanisms themselves, rather than simple distribution changes.
Similarly, Inwoo Hwang et al. use vector quantization to categorize state-action pairs into distinct contexts, allowing reinforcement learning agents to learn localized causal graphs. By relaxing the strict global invariance constraint, frameworks like the one proposed by Moulik Choraria et al. allow models to learn features that remain stable within specific partitions of data, preserving predictive power even when the broader environment shifts.
Continuous and Online Graph Adaptation
When data arrives in continuous streams, systems cannot afford to stop and retrain from scratch. In piecewise-stationary environments—where relationships shift abruptly due to systemic shocks or policy changes—agents must recover causal structures on the fly from streaming feedback (Behzad Nourani-Koliji et al.; Piecewise-Stationary Combinatorial Semi-Bandit with Causally Related Rewards).
To achieve this, researchers have integrated causal inference with online learning and continual learning:
- Dynamic Priors: Virginia Aglietti et al. employ dynamic priors that update expectations based on temporal evolution, allowing sequential decision-making frameworks to maintain accuracy in non-stationary settings.
- Mitigating Catastrophic Forgetting: To prevent models from discarding valid knowledge when data distributions shift, Zhixuan Chu et al. introduce a framework that combines feature distillation and transformation to estimate causal effects from streaming data without needing raw historical data.
- Continual Neural Updates: As argued by Florian Peter Busch et al., merging continual learning with neural causal models allows systems to efficiently update their graphs to incorporate new mechanisms while retaining previously learned, still-valid causal knowledge.