Update app.py

app.py

@@ -6,7 +6,6 @@ import pandas as pd
 import gradio as gr
 import matplotlib.pyplot as plt
 
-# Optional: helps reduce tiny resize flicker in some HF setups
 plt.rcParams["figure.dpi"] = 100
 
 # =========================
@@ -26,9 +25,7 @@ FEATURES = [
 ]
 
 # =========================
-# Cluster anchors (from SPSS)
-# Start state = Cluster 3 (at-risk profile)
-# Target state = Cluster 1 (stable profile)
+# Cluster anchors
 # =========================
 CLUSTER_1 = {
     "Voice": 4.84,
@@ -53,15 +50,11 @@ CLUSTER_3 = {
 VISIBLE_DRIVERS = ["Engagement", "SupportiveGM", "WellBeing", "WorkEnvironment"]
 VISIBLE_LABELS = ["Engagement", "Supportive GM", "Well-Being", "Work Environment"]
 
-
 # =========================
-# SHAP setup (scenario-level)
-# Shows which features drive the current prediction.
-# If SHAP isn't available, we fall back to model feature importance (if available).
+# SHAP setup (optional)
 # =========================
 SHAP_AVAILABLE = False
 explainer = None
-
 try:
     import shap  # noqa: F401
     from shap import TreeExplainer  # type: ignore
@@ -97,12 +90,11 @@ def risk_label(p):
 
 
 def stable_threshold():
-    # threshold line = minimum of the 4 visible drivers in the stable (Cluster 1) profile
     return min(CLUSTER_1[v] for v in VISIBLE_DRIVERS)
 
 
 # =========================
-# Plot: driver bars + threshold
+# Plot: drivers vs threshold
 # =========================
 def make_driver_plot(Engagement, SupportiveGM, WellBeing, WorkEnvironment):
     th = stable_threshold()
@@ -115,7 +107,7 @@ def make_driver_plot(Engagement, SupportiveGM, WellBeing, WorkEnvironment):
     ax.axhline(th, linestyle="--", linewidth=2)
     ax.text(3.05, th, "Stable threshold", va="center")
 
-    ax.set_ylim(1, 5.4)  # extra space above 5
+    ax.set_ylim(1, 5.4)
     ax.set_yticks([1, 2, 3, 4, 5])
     ax.set_ylabel("Survey Score (1–5)")
     ax.set_title("Key Drivers vs Stable Threshold")
@@ -127,54 +119,34 @@ def make_driver_plot(Engagement, SupportiveGM, WellBeing, WorkEnvironment):
 
 
 # =========================
-# Plot: SHAP (or fallback importance)
+# Plot: SHAP (or fallback)
 # =========================
 def make_shap_plot(X: pd.DataFrame):
-    # We’ll show feature impact for the CURRENT prediction (one-row SHAP bar chart).
-    # Exclude ManagementLevel from the display because you don't want mgmt info in the story.
-    display_features = [f for f in FEATURES if f != "ManagementLevel"]
-
     fig, ax = plt.subplots(figsize=(10.5, 4.8))
 
     if SHAP_AVAILABLE and explainer is not None:
         shap_vals = explainer.shap_values(X)
 
-        # shap_values formats vary by model:
-        # - array (n, p)
-        # - list of arrays for classes
-        # We'll pick the "At Risk" class (label 1) if it's a list.
         if isinstance(shap_vals, list):
-            # classes aligned with model.classes_
             classes = list(model.classes_)
             idx = classes.index(1)
             sv = shap_vals[idx][0]
         else:
             sv = shap_vals[0]
 
-        # Build series aligned to columns
-        s = pd.Series(sv, index=X.columns)
-
-        # Drop ManagementLevel for display
-        s = s.drop(labels=["ManagementLevel"], errors="ignore")
-
-        # Rank by absolute contribution
-        s = s.reindex(s.abs().sort_values(ascending=False).index)
-
-        # Plot top 8 (or fewer)
-        top = s.head(8)
-        ax.barh(top.index[::-1], top.values[::-1])
+        s = pd.Series(sv, index=X.columns).drop(labels=["ManagementLevel"], errors="ignore")
+        s = s.reindex(s.abs().sort_values(ascending=False).index).head(8)
 
+        ax.barh(s.index[::-1], s.values[::-1])
         ax.set_title("What drives this prediction (SHAP impact)")
         ax.set_xlabel("Impact on model output (signed)")
         plt.tight_layout()
         return fig
 
-    # ---- Fallback: model feature importance (global) ----
+    # fallback feature importance
     imp = None
-    # sklearn-style
     if hasattr(model, "feature_importances_"):
         imp = pd.Series(model.feature_importances_, index=FEATURES)
-    # CatBoost-style (sometimes)
     elif hasattr(model, "get_feature_importance"):
         try:
             imp = pd.Series(model.get_feature_importance(), index=FEATURES)
@@ -184,17 +156,14 @@ def make_shap_plot(X: pd.DataFrame):
     if imp is None:
         ax.text(
             0.5, 0.5,
-            "SHAP not available in this Space.\nInstall 'shap' to show a SHAP chart.",
+            "SHAP not available.\nAdd 'shap' to requirements.txt for SHAP charts.",
             ha="center", va="center"
         )
         ax.set_axis_off()
         plt.tight_layout()
         return fig
 
-    # Drop ManagementLevel for display
-    imp = imp.drop(labels=["ManagementLevel"], errors="ignore")
-    imp = imp.sort_values(ascending=True).tail(8)
-
+    imp = imp.drop(labels=["ManagementLevel"], errors="ignore").sort_values(ascending=True).tail(8)
     ax.barh(imp.index, imp.values)
     ax.set_title("Feature importance (fallback — not SHAP)")
     ax.set_xlabel("Importance")
@@ -203,21 +172,18 @@ def make_shap_plot(X: pd.DataFrame):
 
 
 # =========================
-# Core prediction
+# Prediction
 # =========================
 def predict(Engagement, SupportiveGM, WellBeing, WorkEnvironment):
-    # visible
     Engagement = clamp(Engagement)
     SupportiveGM = clamp(SupportiveGM)
     WellBeing = clamp(WellBeing)
     WorkEnvironment = clamp(WorkEnvironment)
 
-    # IMPORTANT: model still needs hidden vars. We'll hold them at the stable (Cluster 1) levels.
-    # This keeps the story focused on the 4 drivers you’re showing.
     vals = {
         "Engagement": Engagement,
         "SupportiveGM": SupportiveGM,
-        "ManagementLevel": 2,  # fixed constant; not shown anywhere
+        "ManagementLevel": 2,  # fixed constant; not shown
         "WellBeing": WellBeing,
         "Voice": CLUSTER_1["Voice"],
         "DecisionAutonomy": CLUSTER_1["DecisionAutonomy"],
@@ -227,19 +193,11 @@ def predict(Engagement, SupportiveGM, WellBeing, WorkEnvironment):
 
     X = build_X(vals)
     p = prob_at_risk(X)
-    label = risk_label(p)
-
-    headline = f"Predicted Status: **{label}**"
+    headline = f"Predicted Status: **{risk_label(p)}**"
 
-    driver_fig = make_driver_plot(Engagement, SupportiveGM, WellBeing, WorkEnvironment)
-    shap_fig = make_shap_plot(X)
+    return headline, make_driver_plot(Engagement, SupportiveGM, WellBeing, WorkEnvironment), make_shap_plot(X)
 
-    return headline, driver_fig, shap_fig
 
-
-# =========================
-# Apply recommendation = move to Cluster 1 targets
-# =========================
 def apply_recommendation():
     e = CLUSTER_1["Engagement"]
     s = CLUSTER_1["SupportiveGM"]
@@ -251,17 +209,21 @@ def apply_recommendation():
 
 
 # =========================
-# UI
+# UI (fixed-height plot areas to prevent shaking)
 # =========================
-with gr.Blocks() as demo:
+CSS = """
+.fixed-plot {
+  height: 520px;
+  overflow: hidden;
+}
+"""
+
+with gr.Blocks(css=CSS) as demo:
     gr.Markdown("# Retention Recommendation Simulator")
-    gr.Markdown(
-        "Use the sliders to simulate workplace conditions. "
-        "Click **Apply Recommendation Plan** to move the profile to the stable target."
-    )
+    gr.Markdown("Use the sliders, then click **Predict**. Click **Apply Recommendation Plan** to move to the stable target.")
 
     with gr.Row():
-        with gr.Column():
+        with gr.Column(scale=1):
             Engagement = gr.Slider(1, 5, value=CLUSTER_3["Engagement"], step=0.01, label="Engagement")
             SupportiveGM = gr.Slider(1, 5, value=CLUSTER_3["SupportiveGM"], step=0.01, label="Supportive GM")
             WellBeing = gr.Slider(1, 5, value=CLUSTER_3["WellBeing"], step=0.01, label="Well-Being")
@@ -270,10 +232,16 @@ with gr.Blocks() as demo:
             btn_predict = gr.Button("Predict")
             btn_recommend = gr.Button("Apply Recommendation Plan")
 
-        with gr.Column():
+        with gr.Column(scale=1):
             headline = gr.Markdown()
+
+            gr.HTML('<div class="fixed-plot">')
             driver_plot = gr.Plot(label="Drivers vs Threshold")
+            gr.HTML('</div>')
+
+            gr.HTML('<div class="fixed-plot">')
             shap_plot = gr.Plot(label="SHAP / Feature Impact")
+            gr.HTML('</div>')
 
     btn_predict.click(
         fn=predict,