feat: add alternative neutralization groups (85+ under-explored keys with AC≤100)

alpha_factory/data/brain_groups.py

@@ -0,0 +1,58 @@
+"""
+BRAIN Alternative Neutralization Groups — Under-explored grouping keys.
+Switching from subindustry (AC=272,958) to these (AC≤100) lowers correlation to global pool.
+"""
+from dataclasses import dataclass
+import random
+
+
+@dataclass
+class BrainGroup:
+    id: str
+    coverage: float
+    alpha_count: int
+    description: str
+
+
+STANDARD_GROUPS = [
+    BrainGroup("subindustry", 1.00, 272958, "Subindustry grouping"),
+    BrainGroup("industry", 1.00, 229917, "Industry grouping"),
+    BrainGroup("sector", 1.00, 333086, "Sector grouping"),
+]
+
+ALT_GROUPS = [
+    BrainGroup("pv13_rcsed_6l", 1.00, 3, "Supply chain 6-level hierarchy"),
+    BrainGroup("pv13_di_6l", 1.00, 6, "Direct industry 6-level"),
+    BrainGroup("pv13_hierarchy_min22_513_sector", 0.69, 7, "Hierarchy min22 sector"),
+    BrainGroup("pv13_di_5l", 1.00, 10, "Direct industry 5-level"),
+    BrainGroup("pv13_rha2_min2_513_sector", 0.70, 11, "Revenue hierarchy min2"),
+    BrainGroup("pv13_hierarchy_min51_f1_513_sector", 0.99, 12, "Hierarchy min51 f1"),
+    BrainGroup("pv13_hierarchy_min20_513_sector", 0.69, 12, "Hierarchy min20"),
+    BrainGroup("pv13_rha2_min10_513_sector", 0.70, 13, "Revenue hierarchy min10"),
+    BrainGroup("pv13_hierarchy_min100_2000_513_sector", 0.69, 14, "Hierarchy min100 2k"),
+    BrainGroup("pv13_hierarchy_f3_513_sector", 0.99, 15, "Hierarchy f3"),
+    BrainGroup("pv13_hierarchy_min51_f2_513_sector", 0.99, 16, "Hierarchy min51 f2"),
+    BrainGroup("pv13_hierarchy_min30_3000_mapped_513_sector", 0.99, 17, "Hierarchy min30 3k mapped"),
+    BrainGroup("pv13_rha2_min5_3000_513_sector", 0.99, 18, "Revenue hierarchy min5 3k"),
+    BrainGroup("pv13_hierarchy_min51_f3_513_sector", 0.99, 18, "Hierarchy min51 f3"),
+    BrainGroup("pv13_hierarchys32_513_sector", 0.99, 20, "Hierarchys32"),
+]
+
+PRODUCTION_GROUPS = [g for g in ALT_GROUPS if g.coverage >= 0.90 and g.alpha_count <= 30]
+
+
+def pick_group(min_coverage=0.90, max_ac=50):
+    candidates = [g for g in ALT_GROUPS if g.coverage >= min_coverage and g.alpha_count <= max_ac]
+    if not candidates:
+        return BrainGroup("industry", 1.00, 229917, "Industry grouping")
+    weights = [1.0 / (g.alpha_count + 1) for g in candidates]
+    total = sum(weights)
+    weights = [w / total for w in weights]
+    return random.choices(candidates, weights=weights, k=1)[0]
+
+
+def get_group_for_expression(prefer_novel=True):
+    if prefer_novel:
+        group = pick_group(min_coverage=0.95, max_ac=30)
+        return group.id
+    return "subindustry"