Counterfactual Temporal Trajectory

The defining constraint of counterfactual generation is finding the smallest possible change to the original time series that flips the model's prediction. This is typically formulated as an optimization problem minimizing a distance metric—such as Dynamic Time Warping (DTW) or L1/L2 norm—between the original and counterfactual sequences. The goal is to identify the most efficient intervention point, ensuring the alternative trajectory remains realistic and actionable rather than an arbitrary adversarial example.

Counterfactual trajectories operationalize algorithmic recourse for time-series models. For a financial model predicting loan default, a counterfactual might show how a slightly altered cash-flow sequence would have resulted in approval. Key properties include:

• Actionability: Changes must correspond to features a user can influence
• Proximity: The counterfactual should be as close to the original as possible
• Sparsity: Only a minimal number of time steps should be altered
• Plausibility: The generated path must respect the underlying data distribution

Generating valid counterfactual trajectories involves navigating a complex, non-convex search space. Common techniques include:

• Gradient-based methods: Computing the gradient of the model's decision boundary with respect to the input sequence to iteratively perturb it toward the target class
• Genetic algorithms: Evolving a population of candidate trajectories using crossover and mutation operators guided by a fitness function balancing proximity and prediction flip
• Variational autoencoders (VAEs): Learning a structured latent space of plausible time series and searching within it for counterfactuals, ensuring distributional realism
• Adversarial training: Using a discriminator network to enforce that generated trajectories remain indistinguishable from real data

A raw counterfactual that flips a prediction but violates temporal logic is useless. Plausibility constraints ensure the generated trajectory respects:

• Autocorrelation structure: The counterfactual must preserve the serial dependence patterns of the original domain
• Feature interdependencies: Changes to one variable at a time step must propagate consistently to correlated variables
• Physical or business constraints: E.g., inventory cannot go negative, temperature cannot change instantaneously
• Causal consistency: The counterfactual should not violate known causal relationships encoded in a structural causal model of the time-series system

Assessing the quality of generated counterfactual trajectories requires multiple dimensions:

• Validity: The proportion of counterfactuals that successfully change the model's prediction to the target outcome
• Proximity: The average distance (e.g., Euclidean, DTW) between original and counterfactual sequences
• Sparsity: The number of time steps modified, measured as L0 norm of the perturbation vector
• Plausibility: Measured by the likelihood of the counterfactual under the training data distribution or via discriminator rejection rate
• Diversity: The range of distinct counterfactuals generated, ensuring multiple recourse paths are available

While both counterfactuals and adversarial examples involve perturbing inputs to change predictions, they serve fundamentally different purposes:

• Counterfactuals seek minimal, interpretable, and actionable changes that explain decision boundaries and provide recourse. They must be plausible under the data distribution.
• Adversarial examples seek imperceptible changes designed to expose model vulnerabilities, often using high-frequency noise invisible to humans. They need not be plausible or actionable.
• Counterfactuals are tools for transparency and fairness; adversarial examples are tools for security and robustness testing.

The algorithmic process of creating a realistic, alternative time series that is minimally different from the original but leads to a different model classification or forecast. This is the direct parent concept, focusing on the search for the smallest perturbation vector in the input space.

• Often formulated as a constrained optimization problem
• Balances proximity (small change) with validity (flipped outcome)
• Can use gradient descent in the latent space of autoencoders

A method that adapts Shapley value calculations to assign importance scores to individual time steps in a sequence model's prediction. It provides a game-theoretic foundation for understanding which past observations drive a forecast.

• Distributes credit fairly among lagged features
• Explains the magnitude and direction of a time step's influence
• Computationally intensive, often requiring sampling approximations

The process of identifying which past time steps and features are the actual causal drivers of a model's forecast, often using structural causal models or intervention analysis. This goes beyond correlation to answer 'what if' questions.

• Uses frameworks like Granger causality or PCMCI
• Distinguishes true causes from confounding signals
• Essential for validating counterfactual plausibility

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. It validates that a generated counterfactual trajectory is not an adversarial artifact.

• Tests if explanations change discontinuously
• Measures local Lipschitz continuity of attributions
• Helps ensure recourse actions are reliable

A representation learning approach that separates a model's latent space into independent factors corresponding to static and dynamic attributes, enabling attribution to time-invariant or time-varying concepts. Counterfactuals can be generated by manipulating only the dynamic latent codes.

• Separates content from motion in sequences
• Uses variational autoencoders with KL divergence penalties
• Produces more semantically meaningful alternative trajectories

A quantitative evaluation score that measures how accurately a temporal explanation reflects the true reasoning process of the underlying model by testing its correlation with model behavior under perturbation. It is the gold standard for validating counterfactual trajectories.

• Compares explanation rankings to ablation outcomes
• Uses metrics like area over the perturbation curve (AOPC)
• Ensures the generated trajectory targets genuinely influential steps