Group Methods¶

Group counterfactual methods generate explanations for clusters or subgroups of similar instances. They balance individual precision with broader applicability.

Available Methods¶

Method	Description	Key Feature
ReViCE	Regional variant of PPCEF	Flow-based group CFs
GLANCE	Group-level anchor counterfactuals	Anchor-based grouping
Group GLOBE-CE	Group variant of GLOBE-CE	Subpopulation focus

When to Use Group Methods¶

Group methods are ideal when you need to:

Provide explanations for similar users
Balance personalization with scalability
Identify subpopulation-specific patterns
Generate semi-personalized recourse

How Groups Are Formed¶

Group methods typically cluster instances based on:

Feature similarity
Prediction confidence
Demographic attributes
Custom grouping criteria

flowchart LR
    A[Dataset] --> B[Clustering]
    B --> C[Group 1]
    B --> D[Group 2]
    B --> E[Group N]
    C --> F[Group CF 1]
    D --> G[Group CF 2]
    E --> H[Group CF N]

Example Usage¶

from counterfactuals.cf_methods.group_methods import GroupPPCEF

# Initialize method
method = GroupPPCEF(
    gen_model=flow_model,
    disc_model=classifier,
    disc_model_criterion=criterion,
    device="cuda"
)

# Generate group counterfactuals
result = method.explain(
    X=X_test,
    y_origin=y_test,
    y_target=target_class,
    X_train=X_train,
    y_train=y_train,
    n_groups=5  # Number of groups
)

# Each instance is assigned to a group
print(f"Group assignments: {result.cf_group_ids}")