Change Point Detection Attribution

Unlike general temporal attribution, this method specifically isolates points where the data-generating process changes. It detects shifts in mean, variance, or autocorrelation structure—not just high-magnitude values. A sudden drop in a sensor's variance from 0.8 to 0.1 is flagged as a regime change, and the model's subsequent reliance on this new stable state is quantified. This distinguishes structural breaks from transient noise.

The core output is a score measuring how much a detected change point amplifies or redirects the model's forecast trajectory. A single change point can have an outsized effect, causing the model to discard a long-running trend. Attribution is computed via counterfactual intervention:

• Mask the change point and re-forecast
• Measure the divergence in predicted paths
• The divergence magnitude is the influence score

Many implementations leverage Bayesian Online Change Point Detection (BOCPD) to compute the posterior probability of a change at each time step. The attribution layer then maps these probabilities to the forecasting model's internal state updates. A step with a 0.97 change probability that triggers a hidden state reset in an RNN receives near-total attribution for the new prediction regime.

Attribution is validated by demonstrating that the identified change point is not merely correlated with but causally responsible for the prediction shift. This uses synthetic interventions:

• Artificially insert a known distributional shift at time t
• Verify the attribution method correctly isolates t
• Confirm that removing t restores the pre-change forecast This distinguishes causal drivers from spurious correlations.

Change point attribution enforces a sparsity prior—only a small fraction of time steps are genuine regime shifts. The method penalizes diffuse attribution across many points, forcing the explanation to concentrate on the few critical junctures where the underlying system actually changed. This contrasts with gradient-based methods that often produce dense, noisy saliency maps across the entire sequence.

The final explanation decomposes the model's prediction into pre-change and post-change components. For a financial model forecasting volatility, the attribution might show: 'The volatility spike at t=47 contributed 73% of the elevated forecast for t=48-60, overriding the prior low-volatility regime.' This decomposition enables operators to understand exactly when and why a model changed its mind.

Identifies which past time steps and features are the actual causal drivers of a model's forecast, often using structural causal models or intervention analysis. Unlike purely correlational methods, this approach distinguishes between spurious associations and genuine cause-effect relationships in time series. Key techniques include:

• Granger causality testing for predictive causality
• Do-calculus interventions on structural causal models
• Counterfactual reasoning to isolate treatment effects
• PCMCI algorithm for discovering causal graphs from observational time-series data

A perturbation-based method that systematically removes or masks individual time steps from a sequence to measure the resulting change in the model's output. By observing the prediction delta after ablation, practitioners can rank time steps by importance. Common ablation strategies include:

• Zero masking: replacing values with zero
• Mean imputation: substituting with the series mean
• Noise injection: adding Gaussian noise to obscure signal
• Block ablation: removing contiguous segments to test for change point clusters

Slides a masking window across a time series, occluding segments to generate a saliency map showing which temporal intervals are most critical for a prediction. This technique is particularly effective for identifying change point neighborhoods where statistical properties shift. The resulting heatmap visually communicates:

• Temporal regions of high model sensitivity
• Boundary effects around detected change points
• Duration of influence for each regime shift
• Interaction effects between multiple change points

A quantitative evaluation score that measures how accurately a temporal explanation reflects the true reasoning process of the underlying model. Faithfulness is tested by correlating explanation outputs with model behavior under perturbation. Core evaluation approaches:

• Ablation correlation: does removing high-importance steps change predictions proportionally?
• Comprehensiveness: how much does prediction change when top-k steps are removed?
• Sufficiency: can top-k steps alone reproduce the prediction?
• Monotonicity: does importance ranking remain stable under input perturbations?

An interpretable proxy model, such as a decision tree or linear model, trained to approximate the predictions of a complex temporal model on a local or global scale. Surrogates translate opaque change point detection logic into human-readable rules. Design considerations include:

• Local vs. global approximation scope
• Fidelity-interpretability trade-off in model selection
• Temporal feature engineering to capture lagged effects
• Rule extraction to surface change point heuristics learned by the black-box model

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 that may have taught the model to recognize certain change point patterns. Key properties:

• Upweighting: measures parameter change if a training example is upweighted
• Leave-one-out approximation: avoids costly retraining
• Hessian-based computation: uses second-order gradient information
• Influential change point examples: identifies training sequences that shaped the model's change detection behavior