Synthesising Counterfactual Explanations via Label-Conditional Gaussian Mixture Variational Autoencoders

This paper proposes LAPACE, a novel model-agnostic framework that utilizes a Label-conditional Gaussian Mixture Variational Autoencoder (L-GMVAE) to generate robust, diverse, and plausible counterfactual explanations by synthesizing paths from input data to learned class-specific latent centroids.

The Gist

Imagine you apply for a loan, but an algorithm says "No." You want to know: "What small changes could I make to get a 'Yes'?"

In the world of Artificial Intelligence, the answer to that question is called a Counterfactual Explanation (CE). It's like a "recipe for success" showing you exactly what to tweak (e.g., "increase your savings by $500") to flip the decision.

However, current recipes have problems:

1. They might be fake: The advice might suggest a savings amount that doesn't exist in the real world (not "plausible").
2. They might be fragile: If the bank slightly updates their computer model, the advice might suddenly become useless.
3. They might be boring: They often give you just one option, whereas you might want a few different choices.

This paper introduces a new system called LAPACE (pronounced "Lapace") to fix these problems. Here is how it works, using simple analogies.

The Problem: The "Shaky Map"

Imagine the bank's decision-making process is a giant, foggy map.

• The "No" zone is a swamp.
• The "Yes" zone is a sunny hill.
• You are currently stuck in the swamp.

Old methods try to draw a straight line from your spot to the hill. But because the map is foggy and the terrain is weird, they might draw a line that goes through a cliff (an impossible change) or a line that disappears if the fog shifts (the model updates).

The Solution: Building a "Training Gym" (L-GMVAE)

The authors first built a special AI model called L-GMVAE. Think of this as a gym where the AI learns what a "successful applicant" actually looks like.

Instead of just memorizing real people, the gym creates idealized mannequins (called centroids) for the "Yes" group.

• Imagine the "Yes" group isn't just one type of person. Some are young with high savings; others are older with steady jobs.
• The AI learns to build several different mannequins, each representing a different type of successful applicant.
• Crucially, these mannequins are built to be realistic (they look like real people) and stable (they won't change if the weather shifts).

The Magic Trick: The "Path" (LAPACE)

Once the gym has these perfect mannequins, the LAPACE algorithm does something clever.

Instead of just teleporting you to the mannequin, it creates a smooth walkway (a path) from your current spot in the swamp to the mannequin on the hill.

1. Multiple Routes: Since there are several mannequins (different types of success), you get multiple paths. You can choose the path that is shortest (closest to your current self) or the path that is safest (most robust).
2. The Destination is Fixed: No matter which path you take, they all lead to the same solid, realistic mannequin. This means:
   • Robustness: If the bank changes its rules slightly (the fog shifts), the mannequin is so solid that it still counts as a "Yes."
   • Stability: If you change your application by a tiny bit (a typo), the path still leads to the same mannequin. You don't get a totally different, confusing answer.

Why is this better?

• It's Safe: The path stays on the "data manifold." Imagine walking on a paved road rather than jumping over a ravine. The advice given is always something that could actually happen in the real world.
• It's Flexible: You can choose to walk a short distance (small changes to your life) or a long distance (big changes) depending on how much effort you want to put in.
• It's Private: The "mannequins" are synthetic. The system doesn't steal real people's data to give you advice; it builds new, perfect examples from scratch.
• It Listens: If you say, "I can't change my age," the system can adjust the path to respect that rule, ensuring the advice is actually actionable for you.

The Bottom Line

Think of LAPACE as a GPS for life changes.
Old GPSs might tell you to drive through a lake because it's the "shortest" distance, or they might give you a different route every time you turn the car on.

LAPACE builds a network of sturdy bridges leading to several different, realistic "success destinations." It guarantees that no matter how you tweak your starting point or how the map updates, the bridge still holds, and you can choose the route that fits your life best.

Technical Summary

1. Problem Statement

Counterfactual Explanations (CEs) are crucial for providing algorithmic recourse, suggesting minimal changes to an input to achieve a desired outcome. However, existing methods struggle to simultaneously satisfy multiple, often conflicting, desirable properties within a unified, model-agnostic framework. Key challenges include:

• Robustness: CEs must remain valid under model changes (e.g., retraining) and input perturbations (stability for similar inputs).
• Plausibility: Explanations must reside on the data manifold (be realistic).
• Diversity: Users need multiple distinct recourse options.
• Actionability: Constraints on which features can be changed must be respected.

Current approaches often address these in isolation (e.g., nearest-neighbor methods lack diversity; gradient-based methods lack robustness) or rely on heuristic processes that risk exposing sensitive training data.

2. Methodology

The authors propose a two-stage framework: a novel generative model (L-GMVAE) and a generation algorithm (LAPACE).

A. Label-Conditional Gaussian Mixture VAE (L-GMVAE)

This is a generative model designed to learn a structured latent space tailored for recourse.

• Architecture: It adapts the standard Gaussian Mixture VAE (GMVAE) to be label-conditional.
• Latent Structure: The latent space is partitioned into Gaussian clusters based on class labels. For a target class $y$, the model learns a set of $K/L$ Gaussian components (clusters).
• Training Objective: The model is trained on the dataset using the classifier's predicted labels (not ground truth) to ensure faithfulness to the target model $M$. It optimizes an Evidence Lower Bound (ELBO) consisting of:
  1. Reconstruction Loss: Ensures decoded points are realistic.
  2. KL Divergence (Cluster Assignment): Encourages the use of all available clusters for a class (promoting diversity).
  3. KL Divergence (Latent Variable): Regularizes the latent space to be smooth and continuous.
• Centroids: After training, the mean of each Gaussian component serves as a latent centroid. When decoded, these centroids represent diverse, plausible, and robust prototypes for their respective classes.

B. LAPACE (LAtent PAth Counterfactual Explanations)

LAPACE is a model-agnostic algorithm that synthesizes entire paths of counterfactuals rather than single points.

• Mechanism:
  1. Encode the input $x$ into its latent representation $z_x$.
  2. Identify the latent centroids ($z_{c_j}$) corresponding to the target class $y'$.
  3. Perform linear interpolation in the latent space between $z_x$ and each target centroid: $z_{\tau} = (1-\tau)z_x + \tau z_{c_j}$.
  4. Decode the interpolated points back to the input space to generate a path of CEs.
• Robustness Guarantees:
  • Input Perturbation: Since all paths for a target class converge to the same fixed centroids, small changes in the input $x$ result in similar paths ending at the same robust prototype.
  • Model Changes: The centroids are learned to be deep within the data manifold of the target class, making them likely to retain their classification even if the classifier is retrained.
• Actionability Constraints: The framework supports user-defined constraints (e.g., "feature $i$ must not exceed $\alpha$"). During interpolation, if a constraint is violated, a lightweight gradient descent step is applied to the latent vector to correct it before decoding.

3. Key Contributions

1. L-GMVAE: A new generative model that explicitly maps class labels to sets of Gaussian clusters, creating a structured latent space where centroids act as robust, diverse recourse prototypes.
2. LAPACE Algorithm: A model-agnostic method that generates paths of CEs. This allows users to navigate the trade-off between proximity (early points in the path) and plausibility/robustness (later points near the centroids).
3. Unified Robustness: The approach guarantees perfect robustness to input perturbations (due to fixed convergence points) and high robustness to model changes (due to manifold adherence), addressing a gap in existing literature.
4. Constraint Integration: Demonstrates how actionability constraints can be seamlessly integrated via gradient optimization through the decoder without complex retraining.

4. Experimental Results

The authors evaluated LAPACE on four datasets (Heloc, Wine, Adult, Compas) using Random Forest and Neural Network classifiers, comparing against state-of-the-art baselines (NNCE, FACE, RobXCE, DiCE, DRCE).

• Metrics: Validity, Proximity, Plausibility (LOF), Diversity, Robustness to Model Changes, and Robustness to Input Changes.
• Key Findings:
  • Plausibility: LAPACE achieved the best plausibility scores across all datasets, effectively managing the trade-off with proximity.
  • Robustness: LAPACE-Last (the centroid endpoint) achieved 100% robustness to model changes and perfect stability against input perturbations. It outperformed robust baselines in both metrics.
  • Diversity: The method provided diverse options, particularly on mixed-type datasets.
  • Efficiency: LAPACE is computationally efficient (second only to NNCE) because it requires only forward passes after the initial L-GMVAE training, avoiding iterative optimization for every new input.
  • Actionability: In constrained experiments, LAPACE successfully generated valid CEs satisfying constraints (100% success rate on tested datasets), outperforming modified baseline approaches.
  • Synthetic Data Utility: The L-GMVAE demonstrated high utility in generating synthetic data that preserves the distribution of the original data (TSTR protocol), with centroid classification accuracy consistently at 100%.

5. Significance and Impact

This work bridges the gap between generative modeling and counterfactual explanation by providing a principled, unified framework for robust recourse.

• Trustworthiness: By guaranteeing stability against input noise and model retraining, LAPACE increases the trustworthiness of algorithmic recourse for end-users.
• Privacy: Unlike nearest-neighbor methods, LAPACE generates synthetic CEs, avoiding the leakage of sensitive training data points.
• User Agency: The "path" approach empowers users to choose their preferred balance between minimal change (proximity) and high confidence (plausibility/robustness).
• Future Directions: The structured latent space opens avenues for region-based explanations, causal constraint integration, and application to other data modalities.

In summary, the paper presents a significant advancement in XAI by moving beyond single-point explanations to a robust, path-based generation strategy that inherently satisfies the complex requirements of real-world algorithmic recourse.