Temporal Causal Attribution

Encodes domain knowledge as a directed acyclic graph (DAG) representing causal relationships between variables across time lags. The model's attributions are constrained or validated against this graph to ensure they reflect genuine causal mechanisms rather than spurious correlations. This prevents the explanation from highlighting a time step that is merely correlated with the outcome but has no causal influence.

Applies the do-operator from Pearl's causal calculus to simulate interventions on specific time steps. By setting a feature at time t to a fixed value and observing the change in the forecast distribution, the method isolates the direct causal effect of that intervention. This is distinct from observational conditioning, which can introduce selection bias.

Answers retrospective 'what if' questions about a forecast. Given an observed outcome, it computes the minimal change to a past time step's value that would have caused a different prediction. This requires a three-step process: abduction (inferring exogenous noise), action (intervening on the variable), and prediction (computing the new outcome).

Quantifies predictive causality by testing if past values of a time series significantly reduce the forecast error of another series beyond its own history. In attribution, this serves as a statistical validation layer: a time step is only considered causally salient if it passes a Granger test, filtering out attention weights that do not correspond to true temporal precedence and predictive utility.

Explicitly models and adjusts for hidden confounding variables that influence both the attributed feature and the target. Techniques like instrumental variable analysis or front-door adjustment are applied to decompose the total attribution into direct causal effect and spurious association, ensuring the explanation is robust to unobserved common causes.

Uses algorithms like PCMCI or VAR-LiNGAM to infer a causal graph directly from observational time-series data when a pre-specified DAG is unavailable. The discovered graph is then used to constrain a predictive model's attributions, ensuring that only time steps identified as causal parents in the graph are assigned high importance scores.

A statistical attribution method that quantifies predictive causality between time series. It measures how much past values of one series improve the forecast of another, often used to validate whether a model's attention aligns with known causal structures. Unlike pure correlation, Granger causality tests for temporal precedence.

The application of algorithms like PCMCI or VAR-LiNGAM to time-series data to infer a causal graph directly from observational data. This graph is then used to validate or constrain feature and lag attributions from a predictive model, moving beyond correlation to structural causal models.

A generated alternative time-series path with minimal changes from the original input that would cause a forecasting model to produce a different, desired outcome. This technique answers 'what if' questions and is central to algorithmic recourse in temporal settings.

A quantitative evaluation score that measures how accurately a temporal explanation reflects the true reasoning process of the underlying model. It tests correlation with model behavior under perturbation—if an explanation says a time step is important, removing it must change the output predictably.

A robust statistical method that estimates the effect of removing a specific training sequence on a model's parameters and its prediction for a test sequence. This identifies influential training examples and helps debug data poisoning or memorization issues in temporal models.

A representation learning approach that separates a model's latent space into independent factors corresponding to static and dynamic attributes. This enables attribution to time-invariant concepts (e.g., patient genetics) versus time-varying concepts (e.g., medication dosage), providing structured causal insights.