Update app.py

app.py

@@ -360,26 +360,44 @@ def train_and_save(
     # ----- METRICS BLOCK (MISSING) -----
     roc_auc = float(roc_auc_score(y_test, proba))
     fpr, tpr, roc_thresholds = roc_curve(y_test, proba)
-    best_thr, best_val, cls = find_best_threshold(y_test, proba, metric="f1")
-
+    cls_05 = compute_classification_metrics(y_test, proba, threshold=0.5)
+    best_thr, best_val, cls_best = find_best_threshold(y_test, proba, metric="f1")
     
     metrics = {
         "roc_auc": roc_auc,
         "n_train": int(len(X_train)),
         "n_test": int(len(X_test)),
+    
+        # reference @0.5
+        "threshold@0.5": 0.5,
+        "accuracy@0.5": cls_05["accuracy"],
+        "balanced_accuracy@0.5": cls_05["balanced_accuracy"],
+        "precision@0.5": cls_05["precision"],
+        "recall@0.5": cls_05["recall"],
+        "f1@0.5": cls_05["f1"],
+        "sensitivity@0.5": cls_05["sensitivity"],
+        "specificity@0.5": cls_05["specificity"],
+        "confusion_matrix@0.5": {
+            "tn": cls_05["tn"], "fp": cls_05["fp"],
+            "fn": cls_05["fn"], "tp": cls_05["tp"],
+        },
+    
+        # primary: best F1 threshold
         "best_threshold_by": "f1",
         "best_threshold": float(best_thr),
-        "accuracy@0.5": cls["accuracy"],
-        "balanced_accuracy@0.5": cls["balanced_accuracy"],
-        "precision@0.5": cls["precision"],
-        "recall@0.5": cls["recall"],
-        "f1@0.5": cls["f1"],
-        "sensitivity@0.5": cls["sensitivity"],
-        "specificity@0.5": cls["specificity"],
-        "confusion_matrix@0.5": {
-            "tn": cls["tn"], "fp": cls["fp"],
-            "fn": cls["fn"], "tp": cls["tp"],
+        "best_f1": float(cls_best["f1"]),
+        "accuracy@best": cls_best["accuracy"],
+        "balanced_accuracy@best": cls_best["balanced_accuracy"],
+        "precision@best": cls_best["precision"],
+        "recall@best": cls_best["recall"],
+        "f1@best": cls_best["f1"],
+        "sensitivity@best": cls_best["sensitivity"],
+        "specificity@best": cls_best["specificity"],
+        "confusion_matrix@best": {
+            "tn": cls_best["tn"], "fp": cls_best["fp"],
+            "fn": cls_best["fn"], "tp": cls_best["tp"],
         },
+    
         "roc_curve": {
             "fpr": [float(x) for x in fpr],
             "tpr": [float(x) for x in tpr],
@@ -389,6 +407,7 @@ def train_and_save(
         "calibration": compute_calibration(y_test, proba, n_bins, cal_strategy),
         "decision_curve": decision_curve_analysis(y_test, proba, np.linspace(0.01, 0.99, dca_points)),
     }
+
         
     
     joblib.dump(pipe, "model.joblib")
@@ -409,7 +428,7 @@ def train_and_save(
             "svd_components": int(svd_components) if use_dimred else None,
             "use_feature_selection": bool(use_feature_selection),
             "l1_C": float(l1_C) if use_feature_selection else None,
-            "selection_method": "SelectFromModel(L1 saga, threshold=mean)" if use_feature_selection else None,
+            "selection_method": "SelectFromModel(L1 saga, threshold=median)" if use_feature_selection else None,
             "note": "If SVD is enabled, SHAP becomes component-level (less interpretable)."
         },
         "positive_class": str(pos_class),
@@ -868,9 +887,11 @@ with tab_train:
                 # Show key metrics at threshold 0.5
                 c1, c2, c3, c4 = st.columns(4)
                 c1.metric("ROC AUC", f"{m['roc_auc']:.3f}")
-                c2.metric("Sensitivity (Recall)", f"{m['sensitivity@0.5']:.3f}")
-                c3.metric("Specificity", f"{m['specificity@0.5']:.3f}")
-                c4.metric("F1", f"{m['f1@0.5']:.3f}")
+                c2.metric("Sensitivity (best F1 thr)", f"{m['sensitivity@best']:.3f}")
+                c3.metric("Specificity (best F1 thr)", f"{m['specificity@best']:.3f}")
+                c4.metric("F1 (best)", f"{m['f1@best']:.3f}")
+                st.caption(f"Best threshold (max F1): {m['best_threshold']:.2f}")
+
                 
                 c5, c6, c7, c8 = st.columns(4)
                 c5.metric("Precision", f"{m['precision@0.5']:.3f}")