One counterfactual for a whole group: DiCE_GI turns coverage and diversity into constraints, changing only 4–7 features on HELOC
Synopsis
Counterfactual explanations for a group of instances no longer have to be generated one by one and then aggregated. Earlier group-level methods mostly merged single-instance explanations into one shared action and handled only binary classifiers. DiCE_GI writes coverage and diversity as explicit constraints and solves them with an augmented Lagrangian method, reaching high coverage at low change cost with a single shared intervention (K=1) on synthetic and real datasets; all experiments are in the binary setting.
Interpretation
A counterfactual explanation can now be solved directly for a group of instances: the user names the instances to explain, a set of acceptable target classes and the number K of interventions, and the method returns a set of shared changes that push most of the group toward the target classes. Earlier group-level explainers mainly targeted binary classifiers and aggregated single-instance explanations into one shared action, leaving the multi-target case unaddressed. On three synthetic datasets and four real ones (Adult Income, German Credit, Law School, HELOC), DiCE_GI is compared with GLANCE, EA, CF-Group and GLOBE-CE on coverage, closeness, sparsity and number of unique actions; all reported results are binary because the available group-level baselines are binary-only.
The method treats small changes as the objective and coverage and diversity as constraints that must hold, rather than summing several losses and tuning coefficients by hand. Gradient-based methods usually merge closeness, sparsity and diversity into one weighted objective, and with many instances these terms pull against each other, requiring manual coefficient tuning without clear guarantees. Coverage uses a differentiable soft validity in place of the non-differentiable hard coverage and penalizes only the h=⌈α·|X_e|⌉ hardest instances, with α set to 1, 0.8 and 0.6 in the experiments; diversity has two layers, one requiring different interventions to specialize in different target classes and one spreading interventions apart in feature space with a determinantal point process, the latter weighted by β=0 for binary tasks.
Perspective
All reported results come from the binary setting, because the available group-level counterfactual baselines are designed for binary classifiers; experiments use a single shared intervention (K=1) and explain logistic regression, XGBoost and a one-hidden-layer neural network. The full multi-target capability still needs validation on benchmarks that support multiple target classes, and the authors list extending to K>1 as future work.
The authors state that plausibility and runtime metrics, as well as a subjective user study, are left to future work, and scalability to larger data has not been tested. The multi-target case cannot be compared against baselines on current benchmarks, so its advantage remains unproven.
