fix(ml): integrate Cleanlab to filter noisy SOC labels and prevent model poisoning

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

services/bot-detector/bot_detector/models.py

@@ -29,6 +29,13 @@ if TORCH_AVAILABLE:
 
 if XGB_AVAILABLE:
     import xgboost as xgb
+    from sklearn.model_selection import cross_val_predict
+
+    try:
+        from cleanlab.filter import find_label_issues
+        CLEANLAB_AVAILABLE = True
+    except ImportError:
+        CLEANLAB_AVAILABLE = False
 
 
 # ─── Caches de modèles ─────────────────────────────────────────────────────
@@ -285,6 +292,35 @@ def load_or_train_xgb(name: str, client, features: list, cycle_id: str):
         return None, None
 
     scale_pos = max(1, int((y == 0).sum() / max((y == 1).sum(), 1)))
+
+    # ── Cleanlab : filtrage des labels SOC bruyants ─────────────────────
+    if CLEANLAB_AVAILABLE:
+        try:
+            quick_model = xgb.XGBClassifier(
+                n_estimators=80, max_depth=4, learning_rate=0.15,
+                eval_metric='logloss', random_state=42, n_jobs=-1,
+                tree_method='hist',
+            )
+            pred_probs = cross_val_predict(
+                quick_model, X, y, cv=3, method='predict_proba',
+            )
+            issues = find_label_issues(
+                labels=y, pred_probs=pred_probs,
+            )
+            noisy_idx = issues[issues['is_label_issue'] == True].index.to_numpy()
+            if len(noisy_idx) > 0:
+                keep = np.ones(len(y), dtype=bool)
+                keep[noisy_idx] = False
+                X, y = X[keep], y[keep]
+                pct = len(noisy_idx) / (len(keep)) * 100
+                log_info(
+                    f"[XGB][{name}] Cleanlab : {len(noisy_idx)}/{len(keep)} "
+                    f"labels bruyants supprimés ({pct:.1f}%)"
+                )
+                scale_pos = max(1, int((y == 0).sum() / max((y == 1).sum(), 1)))
+        except Exception as exc:
+            log_info(f"[XGB][{name}] Cleanlab échoué, labels bruts conservés : {exc}")
+
     model = xgb.XGBClassifier(
         n_estimators=200, max_depth=6, learning_rate=0.1,
         scale_pos_weight=scale_pos, eval_metric='logloss',

services/bot-detector/bot_detector/requirements.txt

@@ -7,5 +7,6 @@ hdbscan>=0.8.38
 isotree>=0.6.1
 torch>=2.0
 xgboost>=2.0
+cleanlab>=2.6
 pyyaml>=6.0
 ja4-common @ file:///app/shared/ja4_common