Gradient-Based Counterfactual

The algorithm computes the gradient of the model's loss function with respect to the input features. This gradient vector indicates the direction of steepest ascent toward the target class. The input is iteratively updated using an optimizer like Adam or SGD to minimize a composite objective that balances flipping the prediction while staying close to the original instance.

• Uses backpropagation to compute ∂L/∂x
• Perturbation is proportional to the gradient magnitude
• Enables direct navigation of the decision boundary

The generation process minimizes a weighted multi-term loss function. The primary term encourages the model to predict the target class, while auxiliary terms enforce counterfactual desiderata. A proximity loss (typically L1 or L2) penalizes large deviations from the original input, and an optional diversity term prevents mode collapse when generating multiple counterfactuals.

• Loss = prediction_loss + λ₁·proximity_loss + λ₂·regularization
• Hyperparameters λ control the trade-off between closeness and validity
• Can incorporate feasibility constraints as differentiable penalty terms

Rather than perturbing raw input features directly, many gradient-based methods operate in a learned latent representation space (e.g., the bottleneck of a variational autoencoder). Gradients are computed with respect to the latent code, and the resulting counterfactual is decoded back to the input space. This approach inherently produces more plausible counterfactuals that lie on the data manifold.

• Avoids adversarial, out-of-distribution instances
• Leverages pretrained generative models (VAEs, GANs)
• Ensures counterfactuals resemble realistic data samples

Unlike model-agnostic methods such as Growing Spheres, gradient-based counterfactuals require full access to the model's internal architecture and weights. This white-box requirement means the technique is only applicable to differentiable models like neural networks, logistic regression, and support vector machines with differentiable kernels. It cannot be used with tree-based ensembles or proprietary black-box APIs.

• Requires gradient computation through the model
• Inapplicable to non-differentiable classifiers
• Provides tighter integration with the decision boundary geometry

Because the gradient is computed per input feature, the resulting counterfactual provides inherent feature-level explanations. The magnitude of the gradient for each feature indicates its importance in flipping the prediction. Features with larger absolute gradient values are modified more aggressively, giving users a clear signal about which attributes are most influential for changing the outcome.

• Gradient magnitudes serve as local feature importance scores
• Enables ranking of actionable features
• Supports integration with saliency map visualization techniques

Gradient-based methods can be extended to respect actionability constraints by projecting gradients onto the subspace of mutable features or by zeroing out gradient components for immutable attributes. However, the raw gradient may suggest changes to features that are causally dependent on others, requiring integration with a Structural Causal Model (SCM) to ensure the counterfactual respects causal ordering and avoids recommending infeasible interventions.

• Immutable feature gradients can be masked during optimization
• Causal constraints require additional projection steps
• Trade-off between unconstrained optimality and real-world feasibility