I built a soft voting model and wanted to calculate the feature contribution for an individual prediction. However, I found discrepancies in the result between the feature contributions calculated by myself and the table in the Explainer Dashboard. I fit the same model and processed data but still the result is different. Is there anyway to solve it ?

This is the feature contributions i get from the explainer dashboard

Below are the codes that I calculate the feature contribution manually using LIME. For now I hardcoded the average_population_effect to be the same with the value in Explainer Dashboard to check on the result but it’s different. I also tried to change the configuration for LIME but not much changes on the result

custom_input = np.array([[-1.1146465042123073, -1.1269069848628173, -0.5204541934529272, -0.38306820131416996, -1.8520200101947943, 0.18595422584978583, -1.181043961181701, -0.9282866963312929]])

explainer = lime.lime_tabular.LimeTabularExplainer(
    training_data=risk_factors_test,  # Scaled data
    feature_names=['sex', 'age', 'hypertension', 'heart_disease', 'ever_married', 'avg_glucose_level', 'bmi', 'smoking_status'],
    class_names=['No Stroke', 'Stroke'],
    mode='classification',
    # kernel_width=kernel_width,
    # feature_selection=feature_selection
)

def predict_proba_fn(X):
    return soft_voting_model.predict_proba(X)

# Explain the custom input prediction
exp = explainer.explain_instance(
    custom_input[0], 
    predict_proba_fn, 
    num_features=8, 
    # num_samples=num_samples
)

# Get the explanation as a dataframe
explanation_list = exp.as_list()
explanation_df = pd.DataFrame(explanation_list, columns=['Reason', 'Effect'])

try:
    explanation_df['Effect'] = explanation_df['Effect'].astype(float)
except ValueError:
    # If 'Effect' values are percentages in string format, convert them to numerical values
    explanation_df['Effect'] = explanation_df['Effect'].str.rstrip('%').astype('float') / 100.0

# The average population 
average_population_effect = 52.21  # Updated value from the explainer dashboard

# Calculate the final prediction 
effect_values = explanation_df['Effect']
contributions_sum = effect_values.sum() * 100  # Convert back to percentage points
final_prediction = average_population_effect + contributions_sum

# This is the result
                                     Reason   Effect
0                     Average of population   52.21%
1        hypertension = -0.5204541934529272  -18.24%
2                  bmi = -1.181043961181701  -12.62%
3   avg_glucose_level = 0.18595422584978583    1.35%
4                 sex = -1.1146465042123073   -1.33%
5                   Other features combined    +0.0%
6                 age = -1.1269069848628173    0.97%
7      heart_disease = -0.38306820131416996    0.87%
8      smoking_status = -0.9282866963312929   -0.33%
9        ever_married = -1.8520200101947943    0.00%
10                         Final prediction   22.88%