Interactive Counterfactual Explanations

• Interactive counterfactual explanations are methods that generate minimal, plausible input modifications under user-specified constraints to change model decisions.
• They integrate human-in-the-loop techniques using frameworks like DECE, ViCE, FCEGAN, and CREDENCE to enforce feature masks, range restrictions, and visualization-guided exploration.
• Empirical studies reveal improvements in prediction validity and usability, enabling actionable insights in domains such as finance, healthcare, and document retrieval.

Interactive counterfactual explanations provide users with actionable “what-if” narratives about machine learning model decisions, allowing them to explore and constrain the ways input modifications affect predictions in real time. The central concept is to generate minimal, plausible changes to an input—under explicit, user-driven constraints—that would alter the model’s output to a desired target, with results interpretable at instance, subgroup, or domain-specific levels. This paradigm contrasts with static, one-shot counterfactuals by foregrounding human-in-the-loop constraint specification, adjustment of mutable features, and visualization and exploration of explanation space across multiple modalities and applications.

1. Formal Foundations and Objective Functions

Interactive counterfactual explanation frameworks operationalize the “what-if” question by solving constrained optimization problems. Given a learned function $f$ (classifier, regressor, or ranker), an instance $x$, and a target outcome $y'$, the aim is to find $x'$ as similar as possible to $x$ such that $f(x') = y'$, subject to user-imposed constraints.

DECE formalizes the objective as: $$\min_{c_1:k} L_{\mathrm{valid}} + \lambda_1 L_{\mathrm{dist}} + \lambda_2 L_{\mathrm{div}}$$ where $L_{\mathrm{valid}}$ enforces decision flipping via a margin-based loss, $L_{\mathrm{dist}}$ penalizes distance from $x$ using a per-feature scaled heterogenous metric, and $L_{\mathrm{div}}$ ensures setwise diversity among $k$ returned counterfactuals (Cheng et al., 2020). ViCE formulates a greedy, discretized variant, minimizing support-size and per-feature magnitude of bin-wise perturbations subject to a post-hoc prediction-constraint, allowing instance-level, feature-constrained search (Gomez et al., 2020). CREDENCE generalizes the objective to document or query perturbations in ranking, seeking minimal edits that cause rank threshold crossings (Rorseth et al., 2023).

Flexible generative approaches (e.g. FCEGAN) introduce user-driven binary masks $m \in \{0,1\}^d$ indicating mutable features, encode user constraints as templates $x_{tmp}$, and use adversarial (WGAN-GP) plus divergence and classifier-alignment losses to ensure validity, proximity, and plausibility (Hellemans et al., 24 Feb 2025).

2. Algorithmic Techniques and User Constraints

Interactive systems implement multiple algorithmic safeguards to enforce actionability, realism, and user-driven flexibility. Key methods include:

• Feature locking, masking, or immutability enforcement: Users specify binary masks or lock toggles per feature; optimization proceeds only in the allowed subspace.
• Range and domain restrictions: Drag-brush intervals, sliders, or hard-coded boundaries project counterfactuals to permitted ranges and legal input domains.
• Minimality and sparsity: Counterfactuals are constrained by support size ($\ell_0$ norm), magnitude (bin moves), or proximity metrics—often re-optimized post-hoc for sparser solutions.
• Discrete validity checks: Rounding and iterative nudges ensure returned solutions are valid (e.g., integer-valued where required, flipping output within a discretized domain).

GAN-based architectures leverage templates and immutable resets, ensuring that only mutable features are acted upon during generation. DECE and ViCE use gradient-masked, projected SGD in feature space and greedy bin optimization, respectively; CREDENCE enumerates minimal perturbation sets in documents or queries based on importance scoring.

3. System Architectures and Visualization Paradigms

Interactive counterfactual explanation systems universally adopt multi-tier architectures and tightly coupled visualization components for exploration:

• DECE: A client-server system, combining a backend CF-engine that generates counterfactuals and a React/D3 frontend supporting table-based subgroup exploration, instance-level views, and interactive setting panels. Subgroup histograms, sparklines, and parallel-coordinate plots facilitate transition between global and individual perspectives (Cheng et al., 2020).
• ViCE: Focused on per-instance views, presenting prediction bars, per-feature density distributions, arrow-based deltas indicating minimal perturbed bin-moves, and lock/sort/density toggles for constraint specification and result comparison (Gomez et al., 2020).
• FCEGAN: Organizes inference around counterfactual templates and dynamically specified masks; output counterfactuals are returned only if they satisfy user-imposed mutability constraints (Hellemans et al., 24 Feb 2025).
• CREDENCE: Adopts an explanation workflow specific to information retrieval, supporting document and query perturbation, instance similarity retrieval, and interactive builder interfaces, all integrated through a React/Material UI frontend and high-throughput backend with neural rankers and topic modeling endpoints (Rorseth et al., 2023).

4. Black-Box and Generative Approaches

Black-box counterfactual explanation leverages model-agnostic techniques when gradients or internals are not accessible. FCEGAN achieves validity by training two discriminators—one for the original class and one for the target—using historical predictions instead of classifier gradients, with counterfactual generation being template-driven over the user-mutable subspace (Hellemans et al., 24 Feb 2025). This approach aligns explanations with empirical class distributions and does not require model retraining, achieving strong validity and user alignment on tabular economic and health datasets.

Contrastively, DECE and ViCE require that the black-box model exposes an inference API (and optionally gradients), applying gradient-masked optimization or greedy search as appropriate. In the document ranking domain, CREDENCE interfaces with monoT5 rankers and Lucene/Pyserini indexes, exposing counterfactuals in terms of document or query content changes (Rorseth et al., 2023).

5. Evaluation Metrics and Empirical Findings

Empirical studies of interactive counterfactual systems emphasize both fidelity (do counterfactuals flip model predictions as required) and usability (do user constraints and visualizations support actionable exploration). FCEGAN demonstrates that introducing interactive templates and mask-driven features increases the valid counterfactual fraction by 10–30 percentage points, often outperforming traditional GAN-based and optimization-based baselines, with only minor trade-offs in diversity (Hellemans et al., 24 Feb 2025). DECE use cases document discovery of model weaknesses (e.g., gender bias in credit models, spurious linkages in medical data), while ViCE's visualization flow surfaces dominant features and exposes infeasibility under stricter user constraints (Gomez et al., 2020). CREDENCE highlights the importance of minimal perturbations for both interpretability (sentence removal and term addition in IR) and model auditing, though notes the absence of large-scale user studies (Rorseth et al., 2023).

6. Use Cases Across Domains

Interactive counterfactual explanations have been demonstrated in a range of practical scenarios:

• Finance and risk assessment: Interactive templates surface the minimal regulatory changes or actionable recourses for loan applicants, with instance- and subgroup-based exploration of fairness and bias (Cheng et al., 2020, Hellemans et al., 24 Feb 2025).
• Medical diagnostics: Subgroup-level counterfactuals expose non-intuitive influences (e.g., diabetic neuropathy signals) and guide actionable changes while accounting for locked and plausible clinical ranges (Cheng et al., 2020).
• Robotics: While not elaborated in detail, methods based on counterfactuals with continuous outputs address infeasibility in physical systems and real-time explanation (Gjærum et al., 2022).
• Document retrieval and ranking: CREDENCE illustrates how users can manipulate document content or query terms to flip a document’s relevance status, diagnosing ranking model behavior and exposing reliance on surface statistics (Rorseth et al., 2023).

7. Limitations and Future Directions

Technical limitations of current interactive counterfactual systems include scalability challenges for greedy or combinatorial search, limited support for categorical or multi-class output (noted in ViCE), and restricted perturbation spaces (e.g., sentence removal or term addition only in CREDENCE) (Gomez et al., 2020, Rorseth et al., 2023). Diversity-validity trade-offs in GAN-based methods and the plausible handling of high-dimensional, non-convex spaces remain open research concerns (Hellemans et al., 24 Feb 2025). Absence of formalized user studies or quantitative evaluation of interpretative efficacy is common. Extending interactive counterfactual frameworks to handle richer data modalities (images, time series, structured data), more nuanced user preferences, and regulatory requirements is an active area of exploration.