pages/2_📈_Visualize_the_Results.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Thu May 25 15:30:56 2023

@author: daliagala
"""

### LIBRARIES ###
import streamlit as st
import numpy as np
import pandas as pd 
import matplotlib.pyplot as plt
import plotly.express as px
import plotly.graph_objects as go
from matplotlib_venn import venn2
from sklearn.metrics import confusion_matrix
from utils import display_proportional, plot_data, run_PCA, create_confusion_matrix_heatmap, plot_conf_rates

### PAGE CONFIG ###
st.set_page_config(page_title='FairTargetSim', page_icon=':robot_face:', layout='wide')

hide_st_style = """
            <style>
            #GithubIcon {visibility: hidden;}
            #MainMenu {visibility: hidden;}
            footer {visibility: hidden;}
            header {visibility: hidden;}
            </style>
            """
st.markdown(hide_st_style, unsafe_allow_html=True)

### DICTIONARIES AND CONSTANTS###

colours_education = {
          'Some high school' : 'indigo',
          'High school diploma / GED' : '#7ae99e',
          'Some college' : '#0a68c9',
          'College degree': '#80c4fa',
          'Professional degree': '#f48508',
          "Master's degree" : '#2fa493',
          'Ph.D.' : '#f2a3a1',
          "Associate's degree" : '#fbcc66',
          'Other' : '#fa322f'
        }

colours_country = {
          'AU' : '#7ae99e',
          'US': '#80c4fa',
          'NZ': '#2fa493',
          'CA' : '#fbcc66'
        }

colours_gender = {
          'f' : '#83C9FF',
          'm': '#0067C9'
        }

characteristic_dict = {
          'gender' : colours_gender,
          'education level' : colours_education,
          'country' : colours_country,
          'age' : 'indigo'
        }

pred_dict = {
          'Model A' : 'Predicted_A',
          'Model B' : 'Predicted_B'
        }

prob_dict = {
          'Model A' : 'Prob_1_A',
          'Model B' : 'Prob_1_B'
        }

model_dict = {
          'Model A' : 'Model_A_label',
          'Model B' : 'Model_B_label'
        }

df_keys_dict = {
          'Divided Visual Attention' :'divided_visual_attention',
          'Forward Memory Span' :'forward_memory_span',
          'Arithmetic Reasoning' : 'arithmetic_problem_solving',
          'Grammatical Reasoning'  : 'logical_reasoning',
          'Go/No go': 'adaptive_behaviour_response_inhibition',
          'Reverse_Memory_Span' : 'reverse_memory_span',
          'Verbal List Learning': 'episodic_verbal_learning',
          'Delayed Verbal List Learning': 'delayed_recall',
          'Digit Symbol Coding': 'abstract_symbol_processing_speed',
          'Trail Making Part A' :'numerical_info_processing_speed',
          'Trail Making Part B': 'numerical_and_lexical_info_processing_speed'
        }

### DEFINE ALL SUB-PAGES AS FUNCTIONS TO CALL ###

def mod_prop(cmA, cmB):
    st.markdown('''This section contains model confusion matrices, which give us a lot of information about how good the models which we produced really are. We obtain the confusion matrices by plotting "Actual" labels on one axis, and "Predicted" labels on the other. For both models, in each square of each confusion matrix, we can see which group of candidates it represents, the number of candidates in this group, and what percentage of all candidates assessed they represent.''')
    st.markdown('''
    - **True Negative (TN)**: candidates predicted to have label "0" whose actual label was "0".
    - **False Positive (FP)**: candidates predicted to have label "1" whose actual label was "0".
    - **False Negative (FN)** candidates predicted to have label "0" whose actual label was "1".
    - **True Positive (TP)**: candidates predicted to have label "1" whose actual label was "1".
    ''')
    
    row1_space1, row1_1, row1_space2, row1_2, row1_space3 = st.columns((0.1, 3, 0.1, 3, 0.1))
    with row1_1:
      create_confusion_matrix_heatmap(cmB, "Model A")
    
    with row1_2:
      model = "B"
      create_confusion_matrix_heatmap(cmB, "Model B")
    
    st.subheader("Model Accuracy Rates")
    st.markdown('''We can also represent each model in terms of its accuracy rates, calculated using the numbers of candidates in each group shown in the confusion matrix:''')
    st.markdown('''
    - True Positive Rate (TPR), also called recall, sensitivity or hit rate, is the probability of a positive test result when the result is indeed positive.
    - True Negative Rate (TNR), or specificity/selectivity, is the probability of a negative test result when the test is indeed negative.
    - Positive Predictive Value (PPV) or Precision is the ratio of truly positive results to all positive results.
    - Negative Predictive Value (NPR) is the ratio of truly negative results to all negative results.
    - False Positive Rate (FPR) or fall-out is the probability of assigning a falsely positive result when the test is negative.
    - False Negative Rate (FNR) or miss rate is the probability that a test with a positive result will falsely be assigned a negative result.
    - False Discovery Rate (FDR) it the ratio of false positive results to total number of positive results.
    ''')
    measures_A = plot_conf_rates(cmA)
    measures_B = plot_conf_rates(cmB)
    fig = go.Figure()
    fig.add_trace(go.Bar(
          x=measures_A["Measure"],
          y=measures_A["Score"],  
          name='Model A rates',
          marker_color='rgb(55, 83, 109)'
      ))
    fig.add_trace(go.Bar(
          x=measures_B["Measure"],
          y=measures_B["Score"],
          name='Model B rates',
          marker_color='rgb(26, 118, 255)'
      ))
    fig.update_layout(
          title='Model measures comparison',
          xaxis_tickfont_size=14,
          xaxis=dict(
              title='Measure'),
          yaxis=dict(
              title='Score',
              titlefont_size=16,
              tickfont_size=14,
          ),
          legend=dict(
              bgcolor='rgba(255, 255, 255, 0)',
              bordercolor='rgba(255, 255, 255, 0)'
          ),
    barmode='group',
    bargap=0.15, # gap between bars of adjacent location coordinates.
    bargroupgap=0.1 # gap between bars of the same location coordinate.
      )
    st.plotly_chart(fig, use_container_width = True)
    if st.checkbox("Show tables"):
        row1_space1, row1_1, row1_space2, row1_2, row1_space3 = st.columns((0.1, 3, 0.1, 3, 0.1))
        # CSS to inject contained in a string
        hide_table_row_index = """
        <style>
        thead tr th:first-child {display:none}
        tbody th {display:none}
        </style> """

        # Inject CSS with Markdown
        st.markdown(hide_table_row_index, unsafe_allow_html=True)
        with row1_1:
            st.markdown("Accuracy rates for model A")
            st.table(measures_A)
        with row1_2:
            st.markdown("Accuracy rates for model B")
            st.table(measures_B)

def model_scores(dataframe):
  st.markdown('''This section displays the distribution of scores assigned to each hypothetical employee, according to the values you set on the sliders, compared with the distribution of protected characteristics. These scores are then used to assign labels to the hypothetical employees, creating two distinct datasets to train the models. In essence, these scores explicitly and numerically mimic the viewpoints of hypothetical hiring managers A and B when deciding who to label as "top employees."''')
  st.markdown('''Just as in the original NCPT dataset analysis, the scores obtained by participants in the cognitive games decline with age. This observation provides insight into the potential issue of ageism inherent in the use of gamified hiring processes.''')
  # Create a selectbox to choose a protected characteristic to explore
  plot_radio = st.selectbox('Characteristic to explore', characteristic_dict.keys())
  row2_space1, row2_1, row2_space2 = st.columns((0.1, 5, 0.1))

  with row2_1:
    data = dataframe[["model_A_scores", "model_B_scores", plot_radio]]

    if plot_radio == "age":
        bins= [18,20,30,40,50,60,70,80,90]
        labels = ['18-20','21-30','31-40','41-50','51-60','61-70','71-80','81-90']
        data['age_bins'] = pd.cut(data['age'], bins=bins, labels=labels, right=False)
        plot_radio = 'age_bins'
    colours = ['rgba(93, 164, 214, 0.5)', 'rgba(255, 144, 14, 0.5)']
    c1, c2, c3, c4, c5 = st.columns((0.1, 3, 0.1, 3, 0.1))
    with c2:
        fig = px.box(data, x = plot_radio, y="model_A_scores", labels={"model_A_scores":"Dataset A Input Scores", 'age_bins':"Age"}, title="Input scores in dataset A")
        fig.update_layout(showlegend=False)
        fig.update_traces(marker_color='rgba(93, 164, 214, 0.5)')
        st.plotly_chart(fig, use_container_width=True)
    with c4:
        fig = px.box(data, x = plot_radio, y="model_B_scores", labels={"model_B_scores":"Dataset B Input Scores", 'age_bins':"Age"}, title="Input scores in dataset B")
        fig.update_traces(marker_color = 'rgba(255, 144, 14, 0.5)')
        fig.update_layout(showlegend=False)
        st.plotly_chart(fig, use_container_width=True)

def PCA_general(full_df, dataframe_PCA):
  st.markdown('''On this page, you can see the distribution of the dataset labels which were assigned based on the scores calculated from the slider values you selected previously. Principal Components Analysis, or PCA, is a technique often used to analyse and subsequently visualize datasets where there are many features per single example. This is the case with the NCPT dataset used in our simulator. Specifically, the battery which we used has 11 features per single example, the example being the player of cognitive games, and, in our metaphor, a hypothetical employee or job candidate. It is impossible to plot 11 dimensions, and PCA allows for the visualisation of multidimensional data, while also preserving as much information as possible.''')
  choice = st.radio("What would you like to explore?", ("PCAs", "Components loading"), horizontal = True)
  pcaA, dfA, labelsA, coeffA, componentsA = run_PCA(dataframe_PCA, 'Model_B_label', 'Model_A_label', 2)
  pcaB, dfB, labelsB, coeffB, componentsB = run_PCA(dataframe_PCA, 'Model_A_label', 'Model_B_label', 2)
  loadings = pcaB.components_.T * np.sqrt(pcaB.explained_variance_)
  total_var = pcaA.explained_variance_ratio_.sum() * 100
  dfA = dfA.rename(columns={'target': 'Dataset A'}).reset_index()
  dfB = dfB.rename(columns={'target': 'Dataset B'}).reset_index()
  df_all = pd.merge(dfA, dfB[['index', 'Dataset B']], on='index', how='left')

  conditions = [
      (df_all['Dataset A'] == 1) & (df_all['Dataset B'] == 0),
      (df_all['Dataset B'] == 1) & (df_all['Dataset A'] == 0),
      (df_all['Dataset A'] == 1) & (df_all['Dataset B'] == 1),
      (df_all['Dataset A'] == 0) & (df_all['Dataset B'] == 0)]
      
  values = ['Selected A', 'Selected B', 'Selected both', 'Not selected']
  df_all['All'] = np.select(conditions, values)

  df_all = df_all.drop(["index"], axis = 1)
  df_all.All=pd.Categorical(df_all.All,categories=['Not selected', 'Selected A', 'Selected B', 'Selected both'])
  df_all=df_all.sort_values('All')

  selections_dict = {0: 'Not selected', 1: 'Selected'}
  df_all = df_all.replace({"Dataset A": selections_dict, "Dataset B": selections_dict})

  color_dict_sel = {'Not selected': '#3366CC', 'Selected': 'grey'}

  if "pca_df" not in st.session_state:
    st.session_state.pca_df = df_all

  if choice == "PCAs":
    c1, c2 = st.columns(2)
    with c1:
      fig = px.scatter(st.session_state.pca_df, 
                      x=st.session_state.pca_df['principal component 1'].astype(str),
                      y=st.session_state.pca_df['principal component 2'].astype(str),
                      title='Dataset A PCA',
                      labels={"x": 'PC 1', "y": 'PC 2'},
                      color=st.session_state.pca_df['Dataset A'],
                      color_discrete_map=color_dict_sel)
      fig.update_traces(marker_size = 8)
      st.plotly_chart(fig, use_container_width=True)
    with c2:
      fig = px.scatter(st.session_state.pca_df, 
                      x=st.session_state.pca_df['principal component 1'].astype(str),
                      y=st.session_state.pca_df['principal component 2'].astype(str),
                      title='Dataset B PCA',
                      labels={"x": 'PC 1', "y": 'PC 2'},
                      color=st.session_state.pca_df['Dataset B'],
                      color_discrete_map=color_dict_sel)
      fig.update_traces(marker_size = 8)
      st.plotly_chart(fig, use_container_width=True)
      
    st.markdown(f'''These plots show the reduction of 11 dimensions (11 subtest results) to 2 dimensions. Total Variance for the data is {total_var:.2f}%. Both of the datasets have the same features, therefore they both have the same total variance. Total variance value indicates what percentage of information has been preserved when the dimensionality was reduced. Note that for both datasets, A and B, different points are labelled "1" or "0". This shows that the two datasets represent the two different target variable definitions which were created by you previously. The plots are interactive - zoom in to explore in detail.''')
      
    pcaA, dfA, labelsA, coeffA, componentsA = run_PCA(dataframe_PCA, 'Model_B_label', 'Model_A_label', 2)
    pcaB, dfB, labelsB, coeffB, componentsB = run_PCA(dataframe_PCA, 'Model_A_label', 'Model_B_label', 2)
    loadings = pcaB.components_.T * np.sqrt(pcaB.explained_variance_)
    total_var = pcaA.explained_variance_ratio_.sum() * 100
    dfA = dfA.rename(columns={'target': 'Dataset A'}).reset_index()
    dfB = dfB.rename(columns={'target': 'Dataset B'}).reset_index()
    df_all = pd.merge(dfA, dfB[['index', 'Dataset B']], on='index', how='left')

    conditions = [
        (df_all['Dataset A'] == 1) & (df_all['Dataset B'] == 0),
        (df_all['Dataset B'] == 1) & (df_all['Dataset A'] == 0),
        (df_all['Dataset A'] == 1) & (df_all['Dataset B'] == 1),
        (df_all['Dataset A'] == 0) & (df_all['Dataset B'] == 0)]
        
    values = ['Selected A', 'Selected B', 'Selected both', 'Not selected']
    df_all['All'] = np.select(conditions, values)

    df_all = df_all.drop(["index"], axis = 1)
    df_all.All=pd.Categorical(df_all.All,categories=['Not selected', 'Selected A', 'Selected B', 'Selected both'])
    df_all=df_all.sort_values('All')

    selections_dict = {0: 'Not selected', 1: 'Selected'}
    df_all = df_all.replace({"Dataset A": selections_dict, "Dataset B": selections_dict})

    if "pca_df" not in st.session_state:
      st.session_state.pca_df = df_all

    fig = px.scatter(st.session_state.pca_df, 
                      x=st.session_state.pca_df['principal component 1'],
                      y=st.session_state.pca_df['principal component 2'],
                      title="PCA with labelled groups",
                      color=st.session_state.pca_df["All"], 
                      width = 800, height = 800,
                      color_discrete_sequence=px.colors.qualitative.Safe,
                      opacity = 0.95)
      
    fig.update_yaxes(
        scaleanchor="x",
        scaleratio=1,
      )
    fig.update_traces(marker_size = 10)
    st.plotly_chart(fig)

  if choice == "Components loading":
      c1,c2 = st.columns(2)
      loadings_df = pd.DataFrame(loadings, columns = ["PC1", "PC2"])
      labels_A_proper = { v:k for k,v in df_keys_dict.items()}
      loadings_df["Features"] = labels_A_proper.values()
      with c1:
          fig = px.bar(loadings_df, x="PC1", y="Features", orientation = 'h')
          st.plotly_chart(fig, use_container_width = True)
      with c2:
          fig = px.bar(loadings_df, x="PC2", y="Features", orientation = 'h')
          st.plotly_chart(fig, use_container_width = True)

    #   fig = go.Figure()
    #   fig.add_trace(go.Bar(
    #       x=loadings_df["PC1"],
    #       y=loadings_df["Features"],  
    #       name='Principal Component 1',
    #       marker_color='rgb(55, 83, 109)',
    #       orientation='h'
    #   ))
    #   fig.add_trace(go.Bar(
    #       x=loadings_df["PC2"],
    #       y=loadings_df["Features"],
    #       name='Principal Component 2',
    #       marker_color='rgb(26, 118, 255)',
    #       orientation='h'
    #   ))
    #   fig.update_layout(
    #       title='Component loadings',
    #       xaxis_tickfont_size=14,
    #       xaxis=dict(
    #           title='Loading value'),
    #       yaxis=dict(
    #           title='Feature',
    #           titlefont_size=16,
    #           tickfont_size=14,
    #       ),
    #       legend=dict(
    #           bgcolor='rgba(255, 255, 255, 0)',
    #           bordercolor='rgba(255, 255, 255, 0)'
    #       ),
    # barmode='group',
    # bargap=0.15, # gap between bars of adjacent location coordinates.
    # bargroupgap=0.1 # gap between bars of the same location coordinate.
    #   )
    #   st.plotly_chart(fig, use_container_width = True)
      
      st.markdown('''On this plot, PCA component loadings can be explored. These facilitate the understanding of how much each variable (which there are 11 of) contributes to a particular principal component. Here, the 11 variables were reduced to 2 components, which are labelled PC1 and PC2. The magnitude of the loading (here displayed as the size of the bar in the bar chart) indicates how strong the relationship between the variable and the component is. Therefore, the higher the bar, the stronger the relationship between that component and that variable. The loading's sign can be positive or negative. This indicates whether the principal component and that variable are positively or negatively correlated. We can see that multiple variables are positively correlated with PC2. Two variables, episodic verbal learning and delayed recall are negatively correlated with both of the components.''')


def model_out(full_df):
  st.markdown('''This section highlights the discrepancies between your two models when presented with the same pool of new, previously unseen candidates to label. Specifically, you'll be investigating the candidates assigned a "1" label by both models. These individuals would be those considered for a job interview or chosen for the role, according to your defined target variable.''')
  # Create a selectbox to choose a protected characteristic to explore
  selectbox = st.selectbox('Characteristic to explore', characteristic_dict.keys())
  representation = st.selectbox("Representation", ("absolute", "proportional"))
  row1_space1, row1_1, row1_space2, row1_2, row1_space3 = st.columns((0.1, 3, 0.1, 3, 0.1))
  with row1_1:
    st.subheader("Candidates selected by model A")

    if representation == "absolute":
      # Select predicted data ==1
      data = full_df.loc[full_df['Predicted_A'] == 1]

      # Use function plot_data to plot selected data
      plot_data(data, selectbox, characteristic_dict[selectbox])
    else:
      display_proportional(full_df, selectbox, 'Predicted_A')
      
  with row1_2:
    st.subheader("Candidates selected by model B")
    
    if representation == "absolute":
      # Select predicted data ==1
      data = full_df.loc[full_df['Predicted_B'] == 1]

      # Use function plot_data to plot selected data
      plot_data(data, selectbox, characteristic_dict[selectbox])

    else:
      display_proportional(full_df, selectbox,'Predicted_B')
      
  
  st.markdown('''In this section, you're comparing the model's selections concerning four protected characteristics: age, gender, education level, and country. You can visualize these differences in two ways: "Absolute" or "Proportional".''')
  st.markdown('''"Absolute" representation gives you the raw numbers or percentages of each characteristic chosen. For instance, if the model labeled 5 female candidates and 5 male candidates as "1", the "Absolute" outcome will display as 50% for both genders."Proportional" representation, on the other hand, shows the percentage of a group selected by the model relative to the total number of that group in the input data. For example, if the model evaluated 100 male candidates and selected 5, you will see a 5% representation. If it evaluated 200 female candidates and selected 5, it will show a 2.5% representation.''')
  st.markdown('''If you encounter empty categories in the "Proportional" view, this indicates that while candidates from these categories were evaluated, none were labeled as "1". Hence, their proportional representation amounts to 0%.''')

def dataframe_out(full_df):
  selectbox_M = st.selectbox('Choose which model output to rank by', pred_dict.keys())

  # Select data
  data = full_df.loc[full_df[pred_dict[selectbox_M]] == 1]
  data = data.sort_values(by = prob_dict[selectbox_M], ascending = False)
  data = data[['Candidate ID','Prob_1_A', 'Prob_1_B', 'Predicted_A', 'Predicted_B']]
  data = data.rename(columns={"Prob_1_A": "Ranking, model A", "Prob_1_B": "Ranking, model B", "Predicted_A": "Predicted label A", "Predicted_B": "Predicted label B"})
  data.index = np.arange(1, len(data) + 1)

  st.table(data.style.background_gradient(subset = ["Ranking, model A", "Ranking, model B"], axis=0, vmin=0.40).highlight_max(color = '#FFCD9B', subset = ["Predicted label A", "Predicted label B"], axis=0))
  
  st.markdown("""In this section, you can review the data for all candidates labeled "1" by the selected model, found at the top of the page. Simultaneously, you can observe the labels assigned to these same candidates by the other model. It's likely that there will be instances where candidates chosen by one model weren't selected by the other. Candidates labeled "1" are highlighted in orange in the "Predicted label A" and "Predicted label B" columns.""")
  st.markdown('''In addition to this, you can see the probability with which each candidate was labeled "1". The intensity of the blue color indicates the candidate's ranking position - a darker blue represents a higher ranking (with 1 being the maximum and 0 the minimum). You may notice that some candidates highly ranked by one model may be ranked significantly lower by the other model.''')
  
def venn_diagram(full_df):
  row2_space1, row2_1, row2_space2, row2_2, row2_space3 = st.columns((0.1, 1, 0.1, 1, 0.1))
  with row2_1:
    fig, ax = plt.subplots()

    list_A = full_df.loc[full_df['Predicted_A'] == 1, 'Candidate ID'].astype(int)
    list_B = full_df.loc[full_df['Predicted_B'] == 1, 'Candidate ID'].astype(int)
    set1 = set(list_A)
    set2 = set(list_B)

    venn2([set1, set2], ('Model A', 'Model B'), ax=ax)
    st.pyplot(fig)

  with row2_2:
    st.markdown('''This Venn Diagram visualizes the number of candidates chosen by both models. It's likely that some candidates will be selected by both models, while others may be chosen by only one model. If we consider Model A as the decision of one hiring manager and Model B as another's, it's easy to see how the selection outcome varies depending on the decision-maker. Some candidates may get the opportunity to be hired, while others might not. This serves as an illustration of the inherent arbitrariness in defining the target variable when dealing with highly subjective outcomes.''')
    st.markdown('''For instance, it's straightforward to define a target variable in a classification problem like distinguishing dragonflies from butterflies, where there's little room for ambiguity. However, defining what makes a 'good' employee is far more challenging due to its subjective nature.''')
    


def model_vis(full_df):
  st.markdown('''In this section, you can visualize the demographics of the different subgroups of the data. Firstly, you can see the demographic characteristics of the candidates who have positive labels ("1") and negative labels ("0") which were assigned based on the scores calculated from the slider values you selected previously. Then, you can visualize the demographic distributions of the data which was used for training and evaluation of the models.''')
  choice = st.radio("**Select desired data:**", ("Positive and negative labels", "Training and evaluation data"), horizontal=True)
  if choice == "Positive and negative labels":
    # Create a selectbox to choose a protected characteristic to explore
    selectbox_Lab = st.selectbox('Label to visualize', ('positive labels', 'negative labels'))

    # Create a selectbox to choose a protected characteristic to explore
    selectbox_Char = st.selectbox('Protected characteristic', characteristic_dict.keys())

    row2_space1, row2_1, row2_space2, row2_2, row2_space3 = st.columns((0.1, 3, 0.1, 3, 0.1))

    with row2_1:
      st.subheader("Dataset A")

      # Select test data
      if selectbox_Lab == 'positive labels':
        data = full_df.loc[full_df['Model_A_label'] == 1]
      else:
        data = full_df.loc[full_df['Model_A_label'] == 0]
      
      # Use function plot_data to plot selected data
      plot_data(data, selectbox_Char, characteristic_dict[selectbox_Char])


    with row2_2:
      st.subheader("Dataset B")

      # Select test data
      if selectbox_Lab == 'positive labels':
        data = full_df.loc[full_df['Model_B_label'] == 1]
      else:
        data = full_df.loc[full_df['Model_B_label'] == 0]

      # Use function plot_data to plot selected data
      plot_data(data, selectbox_Char, characteristic_dict[selectbox_Char])
    st.markdown('''You are visualising the demographic composition of those hypothetical employees who were assigned labels "1" or "0" based on your definitions of the target variables. You might see differences in proportions of genders between the two models for the positive labels, as well as a major difference in the age between the positive and negative labels. Visualising the labels in this manner before training the model can help understand and mitigate differences in demographic representation in the modelling outcomes. Likely, if all candidates labelled "1" were in younger age groups, the candidates selected by the model at the deployment stage will also be in younger age groups. Moreover, target variable definition affects the proportional representation. Having defined two target variables, one can choose the dataset and the model which offers more proportional representation.''')


  if choice == "Training and evaluation data":
    # Create a selectbox to choose a protected characteristic to explore
    selectbox = st.selectbox('Characteristic to explore - training and evaluation data', characteristic_dict.keys())
    row1_space1, row1_1, row1_space2, row1_2, row1_space3 = st.columns((0.1, 1, 0.1, 1, 0.1))
    # Plot training data
    with row1_1:
      st.subheader("Training data")

      # Select train data
      train = full_df.loc[full_df["Predicted_A"] == "train"]

      # Use function plot_data to plot selected data
      plot_data(train, selectbox, characteristic_dict[selectbox])

    # Plot test data

    with row1_2:
      st.subheader("Test data")

      # Select test data
      test = full_df.loc[full_df["Predicted_A"] != "train"]

      # Use function plot_data to plot selected data
      plot_data(test, selectbox, characteristic_dict[selectbox])
  
    st.markdown('''To train a machine learning model, the data has to be split into two different sets. The first set is the training data, which will be used to teach the model the relationships between the input features (11 subtest results) and the corresponding labels ("0" and "1", assigned based on your definitions of target variables and the values you chose for the sliders). The second set is the test data, or evaluation data. It is used to assess the performance of the model. This is the data which is used to plot the confusion matrices and calculate the model metrics which you saw at the bottom of the "Define the target variable" page. This is also the data whose features you can explore in "Modelling outcomes". It is important that the training and testing data are balanced. Here, you can compare the demographic composition of the training and evaluation data. The training and evaluation datasets compositions were the same and contained the same candidates and same features for both models A and B. However, the labels for each dataset were different and based on what you selected in "Define target variable".''')

def filter_for_protected(data):
  st.markdown('''Sometimes, the overall model metrics can be deceptive when it comes to predicting the results for different groups under consideration. Ideally, for our models, the varying model metrics would be similar across different groups, which would indicate that the overall model performance is reflected in how this model performs for a given group. It is often not the case, and it is likely that you will see that models A and B perform differently when it comes to those metrics. Even the same model can have different metrics for different subgroups.''')
  st.markdown('''We can represent each model in terms of its accuracy rates, calculated using the numbers of candidates in each group:''')
  st.markdown('''
    - True Positive Rate (TPR), also called recall, sensitivity or hit rate, is the probability of a positive test result when the result is indeed positive.
    - True Negative Rate (TNR), or specificity/selectivity, is the probability of a negative test result when the test is indeed negative.
    - Positive Predictive Value (PPV) or Precision is the ratio of truly positive results to all positive results.
    - Negative Predictive Value (NPR) is the ratio of truly negative results to all negative results.
    - False Positive Rate (FPR) or fall-out is the probability of assigning a falsely positive result when the test is negative.
    - False Negative Rate (FNR) or miss rate is the probability that a test with a positive result will falsely be assigned a negative result.
    - False Discovery Rate (FDR) it the ratio of false positive results to total number of positive results.
    ''')
  model = st.selectbox('Choose which model outputs to assess', pred_dict.keys())
  test = data.loc[data[pred_dict[model]] != "train"]

  selectbox_Char = st.selectbox('Protected characteristic', characteristic_dict.keys())
  if selectbox_Char == 'age':
    bins= [18,20,30,40,50,60,70,80,91]
    labels = ['18-20','21-30','31-40','41-50','51-60','61-70','71-80','81-90']
    test['age_bins'] = pd.cut(test['age'], bins=bins, labels=labels, right=False)
    selectbox_Char = 'age_bins'
  # which_group = st.selectbox('Which group?', test[selectbox_Char].unique())
  df = pd.DataFrame({'Measure': ['True Positive Rate', 'True Negative Rate', 'Positive Predictive Value', 'Negative Predictive Value', 'False Positive Rate', 'False Negative Rate', 'False Discovery Rate']})
  for group in test[selectbox_Char].unique():
      rslt_df = test[test[selectbox_Char] == group] 
      y_true = [int(numeric_string) for numeric_string in rslt_df[model_dict[model]]]
      y_pred = [int(numeric_string) for numeric_string in rslt_df[pred_dict[model]]]
      cm = confusion_matrix(y_true, y_pred)
      if cm.shape == (1,1):
          cm = np.array([[cm[0, 0], 0], [0, 0]])
      d = plot_conf_rates(cm)
      df[f"{group}"] = d["Score"]

  fig = go.Figure()
  for group in test[selectbox_Char].unique():
      fig.add_trace(go.Bar(
          x=df["Measure"],
          y=df[group],  
          name=group
      ))
      
  fig.update_layout(
          title='Model metrics per group',
          xaxis_tickfont_size=14,
          xaxis=dict(
              title='Metric'),
          yaxis=dict(
              title='Score',
              titlefont_size=16,
              tickfont_size=14,
          ),
          legend=dict(
              bgcolor='rgba(255, 255, 255, 0)',
              bordercolor='rgba(255, 255, 255, 0)'
          ),
    barmode='group',
    bargap=0.15, # gap between bars of adjacent location coordinates.
    bargroupgap=0.1 # gap between bars of the same location coordinate.
      )
  st.plotly_chart(fig, use_container_width = True)
  if st.checkbox("Show table of scores"):
    # CSS to inject contained in a string
    hide_table_row_index = """
    <style>
    thead tr th:first-child {display:none}
    tbody th {display:none}
    </style> """

    # Inject CSS with Markdown
    st.markdown(hide_table_row_index, unsafe_allow_html=True)
    st.markdown(f"Accuracy rates for {selectbox_Char}")
    st.table(df)

def data_plot(key1, key2, key3, key4):
  st.title('''Visualize the Results''')
  if key1 not in st.session_state:
    st.error('Cannot train the models if you do not define the target variables. Go to "Define Target Variables"!', icon="🚨")
  else:
    tab1, tab2 = st.tabs(["Demographic", "Non-demographic"])
    with tab1:
      dataframe = st.session_state[key1]
      clean_data = st.session_state[key2]
      st.subheader('''**Select what to explore:**''')
      data_choice = st.radio('''What to explore''', ("Modelling outcomes", "Input data"), horizontal = True, label_visibility = "collapsed")
      if data_choice == "Modelling outcomes":
          st.subheader('''Demographics of the overall modelling outcomes''')
          model_out(dataframe)
          st.subheader('''Demographics of the selected protected groups''')
          filter_for_protected(dataframe)
      else:
          st.subheader('''Demographics of the input scores''')
          model_scores(dataframe)
          st.subheader('''Demographics of the input labels''')
          model_vis(dataframe)
    with tab2:
      dataframe = st.session_state[key1]
      clean_data = st.session_state[key2]
      cmA = st.session_state[key3]
      cmB = st.session_state[key4]
      st.subheader('''**Select what to explore:**''')
      data_choice = st.radio('''Select what to explore:''', ("Modelling outcomes", "Input data"), horizontal = True, label_visibility = "collapsed")
      if data_choice == "Modelling outcomes":
          st.subheader('''Labelled dataframe''')
          dataframe_out(dataframe)
          st.subheader('''Venn Diagram''')
          venn_diagram(dataframe)
          st.subheader('''Model accuracy metrics''')
          mod_prop(cmA, cmB)
      else:
          st.subheader('''Principal Component Analysis''')
          PCA_general(dataframe, clean_data)
    
data_plot('complete_df', 'clean_df', 'cm_A', 'cm_B')