Glossary

Growing Spheres

A model-agnostic counterfactual generation algorithm that searches for the closest counterfactual by expanding a hypersphere around the query instance until the decision boundary is crossed.

Get in touch Learn more

Stylish WeWork-like workspace with hot desks and document wall, professional searching through enterprise knowledge base on a mounted ultrawide display, warm industrial pendants overhead.

MODEL-AGNOSTIC COUNTERFACTUAL SEARCH

What is Growing Spheres?

A foundational algorithm for generating counterfactual explanations by systematically expanding a hypersphere in the feature space until a decision boundary is crossed.

Growing Spheres is a model-agnostic counterfactual generation algorithm that identifies the closest counterfactual instance by iteratively expanding a hypersphere centered on the query point. The algorithm samples points on the surface of an ever-widening sphere until it finds one that crosses the model's decision boundary, guaranteeing a minimal-change explanation without requiring access to model gradients or internal parameters.

Unlike gradient-based methods, Growing Spheres operates purely through black-box query access, making it suitable for any classifier. The process begins with a small radius and incrementally grows it, sampling uniformly from the sphere's surface. Once a valid counterfactual is found inside the sphere, a local search refines it toward the boundary to minimize counterfactual proximity, balancing computational cost with the guarantee of finding a nearby valid instance.

ALGORITHM MECHANICS

Key Features of Growing Spheres

The Growing Spheres algorithm generates counterfactuals by systematically expanding a hypersphere around the query instance until the decision boundary is crossed, ensuring minimal change.

Model-Agnostic Operation

Growing Spheres treats the target model as a black box, requiring only the ability to query for predictions. It does not rely on gradients, internal weights, or architectural assumptions.

Works with any classifier: neural networks, tree ensembles, SVMs
No access to model internals required
Ideal for auditing proprietary or third-party APIs
Operates purely on input-output pairs

Zero

Model Assumptions

Hypersphere Expansion Search

The algorithm initializes a hypersphere centered on the query instance and iteratively expands its radius. At each step, it samples points on the sphere's surface until a point with the target prediction is found.

Begins with radius = 0 at the query point
Uniformly samples from the sphere surface at each expansion step
Guarantees the first valid counterfactual found is close to optimal
Uses L2 distance as the default proximity metric

Two-Phase Generation Process

Growing Spheres operates in two distinct phases to balance exploration and refinement:

Phase 1 — Expansion: The sphere radius grows until a counterfactual is found, establishing an upper bound on the minimal distance.

Phase 2 — Refinement: A binary search between the last unsuccessful radius and the successful radius narrows the distance, producing a counterfactual that lies precisely on the decision boundary.

Sparse Feature Selection

After identifying a boundary-crossing point, Growing Spheres applies a feature selection step to reduce the number of changed features, improving interpretability.

Uses a greedy algorithm to drop feature changes that do not revert the prediction
Prioritizes explanations with fewer altered dimensions
Balances proximity with sparsity for human comprehension
Outputs a minimal set of actionable changes

Handling Immutable Constraints

The algorithm can be extended to respect feasibility constraints by projecting sampled points onto a valid subspace before querying the model.

Immutable features are held constant at their original values
Categorical features can be handled via one-hot encoding and constrained sampling
Ensures generated counterfactuals respect real-world actionability
Integrates with formal action set specifications

Distance Metric Flexibility

While L2 distance is the default, Growing Spheres supports alternative distance functions to align with domain-specific notions of similarity.

Mahalanobis distance accounts for feature correlations and variances
Weighted L2 allows domain experts to encode feature importance
Custom metrics can reflect cost or difficulty of changing each feature
Metric choice directly impacts the plausibility of generated counterfactuals

METHOD COMPARISON

Growing Spheres vs. Other Counterfactual Methods

A technical comparison of the Growing Spheres algorithm against gradient-based and genetic counterfactual generation methods across key operational dimensions.

Feature	Growing Spheres	Gradient-Based	Genetic/Evolutionary
Model Access Required	Black-box (predictions only)	White-box (gradient access)	Black-box (predictions only)
Handles Non-Differentiable Models
Handles Categorical Features
Generates Sparse Counterfactuals
Distance Metric	L2 (Euclidean)	L1, L2, or custom loss	Configurable fitness function
Proximity Guarantee	Local minimum (first boundary crossing)	Local minimum (gradient descent)	Global search (population-based)
Computational Cost	Moderate (iterative sampling)	Low (direct optimization)	High (generational evolution)
Risk of Adversarial Examples	Low (expanding from instance)	Moderate (loss minimization)	Low (diversity pressure)

GROWING SPHERES EXPLAINED

Frequently Asked Questions

Clear, technical answers to the most common questions about the Growing Spheres algorithm for generating counterfactual explanations in machine learning models.

The Growing Spheres algorithm is a model-agnostic counterfactual generation method that identifies the closest counterfactual instance by iteratively expanding a hypersphere in the feature space around the query point until the decision boundary of the target class is crossed. The process operates in two distinct phases. First, a growing phase performs a binary search on the radius of a hypersphere centered on the factual instance, sampling points on the sphere's surface until at least one point is classified with the desired target label. This establishes the minimal distance to the decision boundary. Second, a shrinking phase refines the candidate by moving from the initial boundary-crossing point back toward the query instance along a straight line, using a bisection method to find the exact point where the decision flips. Because it relies solely on querying the model's prediction function and requires no gradient information, Growing Spheres works on any classifier, including black-box APIs and non-differentiable models like random forests.

Enabling Efficiency, Speed & Accuracy

Intelligent Analysis, Decision & Execution

We build AI systems for teams that need search across company data, workflow automation across tools, or AI features inside products and internal software.

Talk to Us

Search across company data

Give teams answers from docs, tickets, runbooks, and product data with sources and permissions.

Useful when people spend too long searching or get different answers from different systems.

Enterprise searchRAGPermissions

Automate internal workflows

Use AI to route work, draft outputs, trigger actions, and keep approvals and logs in place.

Useful when repetitive work moves across multiple tools and teams.

AI agentsWorkflow automationGovernance

Add AI to products and internal tools

Build assistants, guided actions, or decision support into the software your team or customers already use.

Useful when AI needs to be part of the product, not a separate tool.

AI integrationDecision supportModel routing

COUNTERFACTUAL GENERATION ECOSYSTEM

Related Terms

The Growing Spheres algorithm operates within a broader landscape of counterfactual explanation methods. Understanding these adjacent concepts is critical for designing robust recourse and fairness systems.

Counterfactual Proximity

The fundamental objective that Growing Spheres directly optimizes. It quantifies the distance between the original query instance and the generated counterfactual, typically using L1 or L2 norms. The algorithm's core mechanism—expanding a hypersphere until it intersects the decision boundary—guarantees finding the closest counterfactual in Euclidean space, minimizing this metric by construction.

L2 Norm

Default Distance Metric

Decision Boundary

The hypersurface in feature space that separates prediction classes. Growing Spheres explicitly searches for this boundary by sampling points on the surface of an expanding hypersphere. The algorithm's termination condition is the first sphere that contains at least one point classified as the target outcome, effectively identifying the local geometry of the model's decision frontier.

Model-Agnostic Explainability

Growing Spheres is a model-agnostic method, meaning it treats the predictive model as a black box. It requires only query access to the model's prediction function, making it applicable to any classifier—from logistic regression to deep neural networks. This contrasts with gradient-based counterfactual methods that require differentiable model internals.

Sparse Counterfactual

While Growing Spheres finds the closest counterfactual by distance, the resulting instance may alter many features. A sparse counterfactual changes only a small number of features, improving human interpretability. Post-processing steps are often applied to Growing Spheres outputs to enforce sparsity, trading off minimal distance for actionable simplicity.

Algorithmic Recourse

The practical application of counterfactual explanations. Growing Spheres provides the minimal change vector required to flip a negative decision to a positive one. This vector can be translated into actionable recommendations for end-users—for example, specifying exactly how much to increase income or reduce debt to qualify for a loan, enabling concrete recourse.

Plausible Counterfactual

A counterfactual that lies within the high-density region of the training data distribution. Growing Spheres may generate instances on the decision boundary that are unrealistic outliers. Evaluating plausibility using Mahalanobis distance or density estimation is a critical validation step to ensure the generated explanation represents a feasible real-world scenario, not an adversarial artifact.

About the author

Prasad Kumkar

CEO & MD, Inference Systems

Prasad Kumkar is the CEO & MD of Inference Systems and writes about AI systems architecture, LLM infrastructure, model serving, evaluation, and production deployment. Over 5+ years, he has worked across computer vision models, L5 autonomous vehicle systems, and LLM research, with a focus on taking complex AI ideas into real-world engineering systems.

His work and writing cover AI systems, large language models, AI agents, multimodal systems, autonomous systems, inference optimization, RAG, evaluation, and production AI engineering.

Limited slotsGet a Free AI Consultation

How We Work

Custom AI workflows for your Business

One-fit-all AI don't work for modern businesses. At Inferensys, we aim to understand your business & custom requirements; which we use to define most efficient agentic workflows, the data, and the tools for your business.