Temporal Occlusion Analysis

The core mechanism involves sliding a masking window across the input sequence, one time step at a time. At each position, the window occludes the underlying data—typically by replacing it with zeros, a baseline value, or Gaussian noise. The model's output change is recorded, and the process repeats until every possible temporal segment has been evaluated. This exhaustive sweep ensures that localized dependencies at any scale are captured, from single anomalous spikes to extended seasonal patterns.

The output is a time-series saliency map—a heatmap overlaid on the original sequence where color intensity corresponds to feature importance. Regions where occlusion caused a large drop in prediction probability are highlighted as critical. This visual representation allows engineers to instantly identify decision boundaries in time, such as a specific hour before a machine failure or a particular week in a quarterly forecast, making the model's temporal reasoning transparent and auditable.

The choice of occlusion value significantly impacts the explanation's fidelity:

• Zero Masking: Replaces values with 0; effective for normalized data but can create out-of-distribution artifacts.
• Mean Imputation: Substitutes the segment with the series' global or local mean; preserves basic statistics.
• Noise Injection: Adds Gaussian noise to the segment; tests robustness to local perturbations.
• Baseline Replacement: Uses a predefined reference sequence, such as a normal operating pattern, to measure deviation from expected behavior.

The window length is a critical hyperparameter that determines the granularity of the explanation. A window of size 1 evaluates individual time steps, revealing point-level importance. Larger windows capture collective temporal patterns like trends or cycles but risk averaging out fine-grained effects. Multi-scale occlusion—running the analysis with multiple window sizes and aggregating results—provides a more complete picture of both short-term shocks and long-range dependencies.

For a sequence of length n and window size w, the method requires approximately n - w + 1 forward passes through the model. This linear scaling makes it computationally tractable for most production time-series models, unlike exact Shapley value methods which scale exponentially. For long sequences, stride-based occlusion—moving the window by s steps instead of 1—can accelerate the process at the cost of reduced spatial resolution in the resulting saliency map.

Unlike gradient-based techniques such as Temporal Integrated Gradients or Temporal Grad-CAM, occlusion analysis is model-agnostic—it requires no access to internal gradients or architecture-specific hooks. This makes it applicable to any black-box forecasting system, including ensembles and proprietary APIs. However, it is generally slower than gradient methods and can produce noisier explanations, as the perturbation itself may push inputs into regions where the model was never trained to be reliable.

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. Unlike occlusion analysis, which uses a sliding window, ablation typically tests the impact of deleting single points or entire features.

• Key Distinction: Tests independence of steps rather than contiguous blocks
• Use Case: Identifying if a model relies on a single critical timestamp
• Output: A ranked list of time steps by their marginal importance

A visual representation, typically a heatmap overlaid on the input sequence, that highlights the relative importance of each time step for a specific model prediction. Occlusion analysis is one method for generating such maps.

• Common Techniques: Occlusion, Integrated Gradients, Attention Flow
• Visual Encoding: Color intensity maps to importance magnitude
• Application: Rapid identification of critical temporal regions by human auditors

A quantitative evaluation score that measures how accurately a temporal explanation reflects the true reasoning of the underlying model. It tests correlation between attributed importance and actual model behavior under perturbation.

• Core Principle: If a step is important, removing it must change the output
• Metrics: Area Over Perturbation Curve (AOPC), rank correlation
• Validation: Ensures occlusion maps are not misleading artifacts

A robustness evaluation method that introduces small, controlled noise or distortions to a time series to analyze the stability and continuity of a model's explanations. Complements occlusion by testing sensitivity to subtle changes rather than complete masking.

• Perturbation Types: Gaussian noise, time warping, magnitude scaling
• Goal: Verify that explanations do not change drastically with minor input variations
• Stability Index: Quantifies explanation consistency under perturbation

An explainability method that identifies points where statistical properties shift and quantifies their disproportionate influence on subsequent predictions. These change points often correspond to the boundaries of high-importance occlusion windows.

• Detection Algorithms: PELT, Binary Segmentation, Bayesian Online Change Point
• Attribution: Measures how much a change point alters the forecast trajectory
• Synergy: Change points can define optimal occlusion window boundaries

A gradient-based attribution technique that computes the integral of gradients along a path from a baseline to the input, specifically adapted for time-series data. Unlike occlusion, it requires no perturbation and provides a theoretically grounded importance score.

• Axiom: Satisfies completeness—importance scores sum to the prediction difference
• Baseline Choice: Often a zero signal or global mean for time series
• Comparison: Smoother saliency maps than occlusion, but higher computational cost