Update pages/LLM_Trial_1.py

pages/LLM_Trial_1.py

@@ -386,7 +386,6 @@ inconsistent_records = new_samples_df[
 st.write(f"Number of inconsistent records on synthetic data: {len(inconsistent_records)}")
 st.write(inconsistent_records.head())  # Show a few inconsistent rows
 st.write('Helps assess basic data fidelity by checking structural or logical violations.') 
-#st.write('The generative model successfully learned the essential business logic')
 
 
 # ### Visual Comparison
@@ -394,9 +393,6 @@ st.write('Helps assess basic data fidelity by checking structural or logical vio
 # Compare selected variables using histograms
 st.subheader(f"Univariate distribution comparison: real vs synthetic")
 st.write('Shows how well each individual feature is mimicked by the synthetic data.')
-#st.write('The model captures variables like Exposure, VehValue, ClaimAmount, ClaimOcc, and \
-#ClaimNb reasonably well, showing similar overall shapes and ranges. Meanwhile for the others \
-#show a poor replication.')
 
 compare_real_vs_synthetic(
     real_df=X_train,
@@ -408,8 +404,6 @@ compare_real_vs_synthetic(
 
 st.subheader(f"Correlation matrix comparison: real vs synthetic")
 st.write('Evaluates preservation of feature-to-feature relationships.')
-#st.write('Overall the correlation structure is well-preserved, indicating this synthetic data \
-#generation method maintains feature relationships effectively')
 
 # Compute correlation matrices
 corr_matrix_X_train = X_train.corr()
@@ -441,8 +435,6 @@ st.pyplot(fig)
 # Kolmogorov-Smirnov test
 st.subheader("Kolmogorov–Smirnov Test Results")
 st.write('Quantifies the statistical distance between real and synthetic distributions.')
-#st.write('Five variables (VehAge, VehBody, Gender, ClaimOcc, ClaimNb) pass the KS test \
-#with p ≥ 0.05, demonstrating good distributional similarity.')
 
 results = []
 
@@ -478,9 +470,6 @@ st.dataframe(styled_df)
 
 st.subheader('PCA comparison')
 st.write('Assesses similarity in global variance structure and major latent components.')
-#st.write('The synthetic data points substantially overlap with the real data in the principal component space, \
-#indicating the synthetic generation method successfully captures the main variance structure and multivariate  \
-#relationships present in the original dataset.')
 # Load the saved models
 img = mpimg.imread('./LLM/pca_d1_80.png')
 fig=plt.figure(figsize=(10, 8))
@@ -494,10 +483,6 @@ st.pyplot(fig)
 
 st.subheader('UMAP comparison')
 st.write('Examines nonlinear manifold structure and clustering behavior.')
-#st.write('This visualization shows a strong co-location across all three dimensions \
-#indicating the synthetic data successfully captures the complex, high-dimensional structure \
-#of the real data, preserving both local neighborhoods and global manifold geometry essential \
-#for downstream modeling tasks.')
 img = mpimg.imread('./LLM/umap_d1_80.png')
 fig=plt.figure(figsize=(10, 8))
 plt.imshow(img)
@@ -525,9 +510,6 @@ results_cost_2 = run_glm_cost_analysis(new_samples_df, X_test, is_sampled=True,v
 # --- SHAP Feature Importance for Frequency ---
 st.subheader('SHAP Feature Importance for Frequency Model')
 st.write('Shows whether drivers of frequency predictions remain consistent across datasets.')
-#st.write('This SHAP analysis reveals good model consistency: ClaimOcc (claim occurrence) dominates feature importance \
-#in both real and synthetic datasets, suggesting the model has learned stable, meaningful patterns. However, the relative \
-#importance of VehBody increases substantially in synthetic data compared to real data.')
 # Prepare data for frequency model SHAP
 X_train_freq = X_train.drop(['Exposure', 'ClaimNb', 'ClaimAmount'], axis=1, errors='ignore')
 y_train_freq = X_train['ClaimNb']
@@ -590,8 +572,7 @@ plot_glm_shap_importance(
 # --- SHAP Feature Importance for Severity ---
 st.subheader('SHAP Feature Importance for Severity Model')
 st.write('Assesses stability of model explanations for severity outcomes.')
-#st.write('The severity model shows concerning instability between real and synthetic data: \
-#the top features completely flip, with VehBody most important on real data but VehValue dominating synthetic data.')
+
 # Prepare data for severity model SHAP
 X_train_cost_prep = X_train[X_train['ClaimAmount'] != 0].copy()
 X_test_cost_prep = X_test[X_test['ClaimAmount'] != 0].copy()
@@ -720,14 +701,11 @@ inconsistent_records = new_samples_df[
 st.write(f"Number of inconsistent records on synthetic data: {len(inconsistent_records)}")
 st.write(inconsistent_records.head())  # Show a few inconsistent rows
 st.write('Helps assess basic data fidelity by checking structural or logical violations.')
-#st.write('The generative model replaced the business patterns in a right way')
 
 
 # ### Visual Comparison
 st.subheader('Univariate distribution comparison: real vs synthetic')
 st.write('Shows how well each individual feature is mimicked by the synthetic data.')
-#st.write('The model captures variables like ClaimAmount, ClaimOcc, ClaimNb and Gender in a good manner. \
-#Meanwhile fails for the others.')
 
 # Compare selected variables using histograms
 compare_real_vs_synthetic(
@@ -739,8 +717,6 @@ compare_real_vs_synthetic(
 
 st.subheader('Correlation matrix comparison: real vs synthetic')
 st.write('Evaluates preservation of feature-to-feature relationships.')
-#st.write('The synthetic data nearly perfectly replicates the correlation structure, with identical \
-#values across almost all variable pairs.')
 
 # Compute correlation matrices
 corr_matrix_X_train = X_train.corr()
@@ -773,9 +749,6 @@ st.pyplot(fig)
 # Kolmogorov-Smirnov test
 st.subheader('Kolmogorov–Smirnov Test Results')
 st.write('Quantifies the statistical distance between real and synthetic distributions.')
-#st.write('Only four variables (Gender, ClaimNb, ClaimAmount, ClaimOcc) pass the KS test achieving \
-#a perfect p = 1.0000 or close to it, but these successes are primarily on claim-related variables \
-#while demographic and policy features are poorly reproduced.')
 
 
 results = []
@@ -811,11 +784,6 @@ st.dataframe(styled_df)
 # ### PCA Analysis
 st.subheader('PCA comparison')
 st.write('Assesses similarity in global variance structure and major latent components.')
-#st.write('The synthetic points exhibit nearly identical spread, density, and boundary \
-#characteristics as the real data, with minimal outliers and no visible systematic shifts.')
-# Load the saved models
-#scaler = load('./LLM/scaler_pca_model_d2_llm_80.pkl')
-#pca = load('./LLM/pca_model_d2_llm_80.pkl')
 
 img = mpimg.imread('./LLM/pca_d2_80.png')
 fig=plt.figure(figsize=(10, 8))
@@ -827,9 +795,6 @@ st.pyplot(fig)
 # ### UMAP Analysis
 st.subheader('UMAP comparison')
 st.write('Examines nonlinear manifold structure and clustering behavior.')
-#st.write('The plot shows that synthetic points (red) closely overlap the real data (blue), \
-#indicating the generative process preserves the global structure of the feature space. \
-#Minor deviations appear at the edges, but overall the synthetic dataset replicates key clusters well.')
 img = mpimg.imread('./LLM/umap_d2_80.png')
 fig=plt.figure(figsize=(10, 8))
 plt.imshow(img)
@@ -858,9 +823,7 @@ results_cost_4 = run_glm_cost_analysis(new_samples_df, X_test, is_sampled=True,
 # --- SHAP Feature Importance for Frequency ---
 st.subheader('SHAP Feature Importance for Frequency Model')
 st.write('Shows whether drivers of frequency predictions remain consistent across datasets.')
-#st.write('The frequency model demonstrates excellent stability across real and synthetic datasets: \
-#both show OwnerAge as the dominant predictor followed by VehAge, with nearly identical feature importance \
-#rankings and similar magnitude patterns.')
+
 # Prepare data for frequency model SHAP
 X_train_freq = X_train.drop(['Exposure', 'ClaimNb', 'ClaimAmount'], axis=1, errors='ignore')
 y_train_freq = X_train['ClaimNb']
@@ -923,9 +886,7 @@ plot_glm_shap_importance(
 # --- SHAP Feature Importance for Severity ---
 st.subheader('SHAP Feature Importance for Severity Model')
 st.write('Assesses stability of model explanations for severity outcomes')
-#st.write('The severity model shows strong consistency between real and synthetic data: \
-#VehAge clearly dominates as the primary driver in both datasets, followed by OwnerAge \
-#as a distant second.')
+
 # Prepare data for severity model SHAP
 X_train_cost_prep = X_train[X_train['ClaimAmount'] != 0].copy()
 X_test_cost_prep = X_test[X_test['ClaimAmount'] != 0].copy()