fix(hpo): exclude DOMIAS from the optimization objective

configs/config_hepatitis.yaml

@@ -86,7 +86,7 @@ generation:
       stats: [prdc, alpha_precision, wasserstein_dist, inv_kl_divergence]
       sanity: [nearest_syn_neighbor_distance]
       performance: [xgb]
-      privacy: [identifiability_score, DomiasMIA_prior]
+      privacy: [identifiability_score]
     storage: null
     best_params_path: null
     final_n_iter_override: null

configs/config_loris.yaml

@@ -117,7 +117,7 @@ generation:
       stats: [prdc, alpha_precision, wasserstein_dist, inv_kl_divergence]
       sanity: [nearest_syn_neighbor_distance]
       performance: [xgb]
-      privacy: [identifiability_score, DomiasMIA_prior]
+      privacy: [identifiability_score]
     storage: null
     best_params_path: null
     final_n_iter_override: null

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "synthdata"
-version = "0.7.3"
+version = "0.7.4"
 description = "Sandbox for synthetic data generation and evaluation"
 authors = [
     {name = "Alp Erkent", email = "alp.erkent@childmind.org"},

synthdata/config.py

@@ -293,7 +293,9 @@ class HPOConfig:
             ],
             "sanity": ["nearest_syn_neighbor_distance"],
             "performance": ["xgb"],
-            "privacy": ["identifiability_score", "DomiasMIA_prior"],
+            # DOMIAS is deferred from HPO until its high-dimensional KDE
+            # failure modes have a separately validated treatment.
+            "privacy": ["identifiability_score"],
         }
     )
     #: Optuna storage URL, e.g. "sqlite:///output/dataset/optuna_studies.db".

tests/unit/test_config.py

@@ -32,6 +32,11 @@ def test_empty_dict_returns_defaults(self):
         cfg = _from_dict(Config, {})
         assert cfg == Config()
 
+    def test_default_hpo_privacy_objective_excludes_domias(self):
+        cfg = _from_dict(Config, {})
+
+        assert cfg.generation.hpo.metric_config["privacy"] == ["identifiability_score"]
+
     def test_flat_fields_applied(self):
         cfg = _from_dict(Config, {"name": "mydata", "seed": 7})
         assert cfg.name == "mydata"
@@ -91,6 +96,13 @@ def test_refidiff_benchmark_profiles_load(self, config_name):
         assert cfg.imputation.refidiff.catboost_warmup_iterations == 1000
         assert cfg.imputation.benchmark.enabled
 
+    @pytest.mark.parametrize("config_name", ["config_hepatitis.yaml", "config_loris.yaml"])
+    def test_shipped_hpo_profiles_exclude_domias(self, config_name):
+        root = Path(__file__).parents[2]
+        cfg = load_config(root / "configs" / config_name)
+
+        assert cfg.generation.hpo.metric_config["privacy"] == ["identifiability_score"]
+
     def test_evaluation_binary_target_nested_dict_builds_nested_dataclass(self):
         cfg = _from_dict(
             Config,

tests/unit/test_synthcity_eval.py

@@ -16,6 +16,9 @@ def test_default_selection_uses_native_attack_category(self):
         assert result["attack"] == SYNTHCITY_METRIC_CONFIG["attack"]
         assert "attacks" not in result
 
+    def test_final_evaluation_catalog_retains_domias(self):
+        assert "DomiasMIA_prior" in SYNTHCITY_METRIC_CONFIG["privacy"]
+
     def test_privacy_category_includes_attack_metrics(self):
         result = resolve_metric_config(FrameworkSelectionConfig(categories=["privacy"]))