pages/go_further.py

import os
import re
import time
import streamlit as st
import matplotlib.pyplot as plt
import pandas as pd
import numpy as np
import altair as alt
import plotly.express as px

from st_pages import add_indentation
from utils import load_data_csv

from sklearn.datasets import fetch_california_housing
from sklearn.compose import make_column_selector as selector
from sklearn.model_selection import train_test_split
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import MinMaxScaler, StandardScaler, OneHotEncoder
from sklearn.neighbors import KNeighborsClassifier
from sklearn.tree import DecisionTreeClassifier
from sklearn.ensemble import RandomForestClassifier
from sklearn.compose import ColumnTransformer
from sklearn.metrics import confusion_matrix


st.set_page_config(layout="wide")


#######################################################################################################
#                                          FUNCTIONS 
#######################################################################################################

@st.cache_data(ttl=3600)
def model_training(X, y, model_dict, _num_transformer=MinMaxScaler(), 
                   _cat_transformer=OneHotEncoder()):
    
    model = model_dict["model"]
    param = model_dict["param"]
    explainability = False
    feature_imp = None
    
    if model == "K-nearest-neighbor 🏘️":
        model_sklearn = KNeighborsClassifier(n_neighbors=param)

    if model == "Decision Tree 🌳":
        model_sklearn = DecisionTreeClassifier(max_depth=param)
        explainability = True

    if model == "Random Forest 🏕️":
        model_sklearn = RandomForestClassifier(max_depth=param)
        explainability = True


    X_train, X_test, y_train, y_test = train_test_split(X, y, stratify=y, test_size=0.33)
    preprocessor = ColumnTransformer(
    transformers=[
        ("numerical", _num_transformer, selector(dtype_exclude="category")),
        ("categorical", _cat_transformer, selector(dtype_include="category")),
    ])
    
    pipe = Pipeline(
        steps=[("preprocessor", preprocessor), ("classifier", model_sklearn)])
    pipe.fit(X_train, y_train)
    
    feature_names = pipe[:-1].get_feature_names_out()
    feature_names = [name.split("__")[1] for name in feature_names]
    feature_names = [name.split("_")[0] if "_" in name else name for name in feature_names]
    
    y_pred = pipe.predict(X_test)
  
    clf = pipe[-1]
    cm = confusion_matrix(y_test, y_pred, labels=clf.classes_, normalize='pred')

    if explainability:
        feature_imp = clf.feature_importances_
    
    labels = clf.classes_

    return np.diag(cm), feature_imp, feature_names, labels


def see_code(model):
    if model == "K-nearest-neighbor 🏘️":
        model_sklearn = "KNeighborsClassifier(n_neighbors=6)"

    if model == "Decision Tree 🌳":
        model_sklearn = "DecisionTreeClassifier()"

    if model == "Random Forest 🏕️":
        model_sklearn = "RandomForestClassifier()"
        
    code = f'''# Split data into train and test sets
X_train, X_test, y_train, y_test = train_test_split(X, y, stratify=y, test_size=0.33)

# Build data preprocessing step to numerical and categorical/text variables
preprocessor = ColumnTransformer(
transformers=[
    ("numerical", num_transformer, selector(dtype_exclude="category")),
    ("categorical", cat_transformer, selector(dtype_include="category")),
])

# Train the model with the preprocessing step 
pipe = Pipeline(
    steps=[("preprocessor", preprocessor), ("classifier", {model_sklearn})])
pipe.fit(X_train, y_train)

# Predict values for the test set 
y_pred = pipe.predict(X_test)

# Compute confusion matrix to get the accuracy for each label
clf = pipe[-1]
cm = confusion_matrix(y_test, y_pred, labels=clf.classes_, normalize='pred')
scores = np.diag(cm)
'''

    st.warning("""**Note**: The following code uses functions from popular Python Data Science libraries `numpy` and `scikit-learn`.""")
    st.code(code, language='python')




##############################################################################################
#                                    START OF THE PAGE
##############################################################################################

st.image("images/ML_header.jpg")
st.markdown("# Go further 🚀")
st.markdown("""This page allows you to test and compare the results of different AI models, and gain a deeper understanding of how they function. <br>
            It includes three different types of **classification models** with Python code illustrations, as well as four datasets to choose from.
            """, unsafe_allow_html=True) 

# st.markdown("""**Reminder**: Classification models are AI models that are trained to predict a finite number of values/categories.
#         Examples can be found in the *Supervised vs Unsupervised* page with the credit score classification and customer churn prediction use cases.""")

st.warning("""**Note**: Different types of models exists for most Machine Learning tasks. 
           Models tend to vary in complexity and picking which one to train for a specific use case isn't always straightforward. 
           Complex model might output better results but take longer to make predictions.
           The model selection step requires a good amount of testing by practitioners.""")

st.markdown("""All of the classification models used in this page come from `scikit-learn`, which is a popular Data Science library in Python.""")
try:
    st.link_button("Go to the scikit-learn website", "https://scikit-learn.org/stable/index.html")
except:
    st.markdown("You need internet connexion to access the link.")

st.markdown("  ")
st.divider()


path_data = r'data/other_data'

st.markdown("# Classification ")
st.markdown("""**Reminder**: Classification models are AI models that are trained to predict a finite number of values/categories.
        Examples can be found in the *Supervised vs Unsupervised* page with the credit score classification and customer churn prediction use cases.""")
st.markdown("  ")
st.markdown("  ")

########################## SELECT A DATASET ###############################

st.markdown("### Select a dataset 📋")
st.markdown("""To perform the classification task, you can choose between three different datasets: **Wine quality**, **Titanic** and **Car evaluation**. <br>
            Each dataset will be shown in its original format and will go through pre-processing steps to insure its quality and usability for the chosen model.
            """, unsafe_allow_html=True)

st.warning("""**Note:** The performance of a Machine Learning model is sensitive to the data being used to train it.
    Data cleaning and pre-processing are usually as important as training the AI model. These steps can include removing missing values, identifying outliers and transforming columns from text to numbers.""")

select_data = st.selectbox("Choose an option", ["Wine quality 🍷", "Titanic 🚢", "Car evaluation 🚙", "Diabetes 👩‍⚕️"]) #label_visibility="collapsed")
st.markdown(" ")

if select_data =="Wine quality 🍷":
    # Load data and clean it
    data = load_data_csv(path_data, "winequality.csv")
    data = data.loc[data["residual sugar"] < 40]
    data = data.loc[data["free sulfur dioxide"] < 200]
    data = data.loc[data["total sulfur dioxide"] < 400]
    data.drop(columns=["free sulfur dioxide"], inplace=True)

    X = data.drop(columns=["quality"])
    y = data["quality"]

    # Information on the data
    st.info("""**About the data**: The goal of the wine quality dataset is to **predict the quality** of different wines using their formulation. 
                The target in this use case is the `quality` variable which has two possible values (Good and Mediocre).""")

    # View data
    view_data = st.checkbox("View the data", key="wine")
    if view_data:
        st.dataframe(data)

    
if select_data == "Titanic 🚢":
    # Load data and clean it
    data = load_data_csv(path_data, "titanic.csv")
    data = data.drop(columns=["Name","Cabin","Ticket","PassengerId"]).dropna()
    data["Survived"] = data["Survived"].map({0: "Died", 1:"Survived"})
    data.rename({"Sex":"Gender"}, axis=1, inplace=True)
    data["Age"] = data["Age"].astype(int)
    data["Fare"] = data["Fare"].round(2)
    
    cat_columns = data.select_dtypes(include="object").columns
    data[cat_columns] = data[cat_columns].astype("category")

    X = data.drop(columns=["Survived"])
    y = data["Survived"]

    # Information on the data
    st.info("""**About the data**: The goal of the titanic dataset is to **predict whether a passenger on the ship survived**. 
                The target in this use case is the `Survived` variable which has two possible values (Died or Survived).
                """)

    # View data
    view_data = st.checkbox("View the data", key="titanic")
    if view_data:
        st.dataframe(data)
        
    # About the variables
    about_var = st.checkbox("Information on the variables", key="titanic-var")
    if about_var:
        st.markdown("""
- **Survived**: Survival (Died or Survived)
- **Pclass**: Ticket class of the passenger (1=First, 2=Second, 3=Third)
- **Gender**: Gender
- **Age**: Age in years
- **SibSp**: Number of siblings aboard the Titanic
- **Parch**: Number of parents/children aboard the Titanic
- **Fare**: Passenger fare
- **Embarked**: Port of Embarkation (C=Cherbourg, Q=Queenstown, S=Southampton)""")

if select_data == "Car evaluation 🚙":
    # Load data and clean it
    data = load_data_csv(path_data, "car.csv")
    data.rename({"Price":"Buying"}, axis=1, inplace=True)
    cat_columns = data.select_dtypes(include="object").columns
    data[cat_columns] = data[cat_columns].astype("category")

    X = data.drop(columns="Evaluation")
    y = data["Evaluation"]

    # Information on the data
    st.info("""**About the data**: The goal of the car evaluation dataset is to predict the evaluation made about a car before being sold.
            The target in this use case is the `Evaluation` variable, which has two possible values (Not acceptable or acceptable)""")

    # View data
    view_data = st.checkbox("View the data", key="car")
    if view_data:
        st.dataframe(data)

    # View data
    about_var = st.checkbox("Information on the variables", key="car-var")
    if about_var:
        st.markdown("""
- **Buying**: Buying price of the vehicule (Very high, high, medium, low)
- **Maintenance**: Price for maintenance (Very high, high, medium, low)
- **Doors**: Number of doors in the vehicule (2, 3, 4, 5 or more)
- **Persons**: Capacity in terms of persons to carry (2, 4, more)
- **Luggage boot**: Size of luggage boot 
- **Safety**: Estimated safety of the car (low, medium, high)
- **Evaluation**: Evaluation level (unacceptable, acceptable)""")
        

if select_data == "Diabetes 👩‍⚕️":
    # Load data and clean it
    data = load_data_csv(path_data, "diabetes.csv")
    data["Outcome"] = data["Outcome"].map({1:"Yes", 0:"No"})
    #data.drop(columns=["DiabetesPedigreeFunction"], inplace=True)
    # data.rename({"Price":"Buying"}, axis=1, inplace=True)
    cat_columns = data.select_dtypes(include="object").columns
    data[cat_columns] = data[cat_columns].astype("category")

    X = data.drop(columns="Outcome")
    y = data["Outcome"]


    # Information on the data
    st.info("""**About the data**: The goal of the diabetes dataset is to predict whether a patient has diabetes.
            The target in this use case is the `Outcome` variable, which has two possible values (Yes or No)""")

    # View data
    view_data = st.checkbox("View the data", key="diabetes")
    if view_data:
        st.dataframe(data)

    # View data
    about_var = st.checkbox("Information on the variables", key="car-var")
    if about_var:
        st.markdown("""
- **Pregnancies**: Number of pregnancies had
- **Glucose**: The level of glucose in the patient's blood
- **BloodPressure**: Blood pressure measurement
- **SkinThickness**: Thickness of the skin
- **Insulin**: Level of insulin in the blood
- **BMI**: Body mass index
- **DiabetesPedigreeFunction**: Likelihood of diabetes depending on the patient's age and diabetic family history
- **Age**: Age of the patient
- **Outcome**: Whether the patient has diabetes (Yes or No)""")

st.markdown(" ")
st.markdown(" ")


########################## SELECT A MODEL ###############################

st.markdown("### Select a model 📚")
st.markdown("""You can choose between three types of classification models: **K nearest neighbors (KNN)**, **Decision Trees** and **Random Forests**. <br>
    For each model, you will be given a short explanation as to how they function.
    """, unsafe_allow_html=True)

select_model = st.selectbox("**Choose an option**", ["K-nearest-neighbor 🏘️", "Decision Tree 🌳", "Random Forest 🏕️"])
st.markdown(" ")


if select_model == "K-nearest-neighbor 🏘️":
    #st.markdown("#### Model: K-nearest-neighbor")
    st.info("""**About the model**: K-nearest-neighbor (or KNN) is a type of classification model that uses neighboring points to classify new data.
            When trying to predict a class to new data points, the algorithm will look at points in close proximity (or in its neighborhood) to make a decision.
            The most common class among its neighborhood will then be assigned to the data point.""")
    
    select_param = 6
    model_dict = {"model":select_model, "param":select_param}

    learn_model = st.checkbox("Learn more", key="knn")
    if learn_model:
        st.markdown("""An important parameter in KNN algorithms is the number of points to choose as neighboors. <br>
                The image below shows two cases where the number of neighboors (k) are equal to 3 and 6.
- When k is equal to 3, the most common class is **Classe B**. The red point will then be predicted as Classe B.
- When k is equal to 6, the  the most common class is **Classe A**. The red point will then be predicted as Classe A.""", 
                unsafe_allow_html=True)
        st.image("images/knn.png", width=600)
        st.markdown("""K-nearest-neighbor algorithm are popular for their simplicity. <br>
                        This can be a drawback for use cases/dataset that require a more complex approach to make accurate predictions.""", unsafe_allow_html=True)
    
    see_code_box = st.checkbox("See the code", key='knn_code')
    if see_code_box:
        see_code(select_model)
    

if select_model == "Decision Tree 🌳":
    st.info("""**About the model**: Decision trees are classification model that split the prediction task into a succession of decisions, each with only two possible outcomes.
            These decisions can be visualized as a tree, with data points arriving from the top of the tree and landing at final "prediction regions".""")
    
    select_param = None
    model_dict = {"model":select_model, "param":select_param}
    
    learn_model = st.checkbox("Learn more", key="tree")
    if learn_model:        
        st.markdown("""The following image showcases a decision tree that was built to predict whether a **bank should give out a loan** to a client. <br>
                    The data used to train the model has each client's **age**, **salary** and **number of children**.""", unsafe_allow_html=True)
        
        st.markdown("""To predict whether a client gets a loan, the client's data goes through each 'question' in the tree and **gets assigned the class of the region it fell into**. <br>
                    For example, a client that is under 30 years old and has a lower salary than 2500$ will not be awarded a loan by the model.""", unsafe_allow_html=True)
        
        st.image("images/decisiontree.png", width=800)
        st.markdown("""Decision tree models are popular as they are easy to interpret. <br>
                    The higher the variable is on the tree, the more important it is in the decision process.""", unsafe_allow_html=True)
        
    see_code_box = st.checkbox("See the code", key='tree_code')
    if see_code_box:
        see_code(select_model)
        


if select_model == "Random Forest 🏕️":
    st.info("""**About the model:** Random Forest models generate multiple decision tree models to make predictions. 
            The main drawback of decision trees is that their predictions can be unstable, meaning that their output often changes.
            Random Forest models aggregate the predictions of multiple decision trees to reduce this unstability and improve robustness.""")
    
    select_param = None
    model_dict = {"model":select_model, "param":select_param}

    learn_model = st.checkbox("Learn more", key="tree")
    if learn_model:
        st.markdown("""Random Forests classifiers aggregate results by apply **majority voting**, which means selecting the class that was most often predicted by trees as the final prediction. 
                    In the following image, the random forest model built four decision trees, who each have made their own final prediction. <br>"""
                    , unsafe_allow_html=True)
        
        st.markdown("""Class C was predicted twice, whereas Class B et D where only predicted once. <br> 
                    The final prediction of the random forest model is thus Class C.""", unsafe_allow_html=True)
        
        st.image("images/randomforest.png", width=800)

    see_code_box = st.checkbox("See the code", key='forest_code')
    if see_code_box:
        see_code(select_model)



st.markdown(" ")
st.markdown(" ")

########################## RUN THE MODEL ###############################

st.markdown("### Train the model ⚙️")
st.markdown("""Now, you can build the chosen classification model and use the selected dataset to train it. <br>
            You will get the model's accuracy in predicting each category, as well as the importance of each variable in the final predictions.""", unsafe_allow_html=True)

st.warning("""**Note**: Most machine learning models have an element of randomness in their predictions. 
            This explains why a model's accuracy might change even if you run it with the same dataset.""")

st.markdown(f"""You've selected the **{select_data}** dataset and the **{select_model}** model.""")


run_model = st.button("Run model", type="primary")

if run_model:
    score, feature_imp, feature_names, labels = model_training(X, y, model_dict, _num_transformer=StandardScaler())
    
    if select_model in ["Decision Tree 🌳", "Random Forest 🏕️"]: # show explainability for decision tree, random firest
        tab1, tab2 = st.tabs(["Accuracy", "Explainability"])

        with tab1:
            if select_data == "Diabetes 👩‍⚕️":
                st.error("""**Important**: The Diabetes dataset only contains information on 768 patients. 500 patients don't have diabetes and 268 do have the disease. 
                            This small number of patient data explains why the model's performance isn't optimal. 
                            Additional data collection should be conducted to improve results, as well as hyperparameter tuning (see explanation after graph).""")
                
            score_df = pd.DataFrame({"label":labels, "accuracy":np.round(score*100)})
            fig = px.bar(score_df, x="label", y="accuracy", color="label", title="Accuracy results", text_auto=True)
            st.plotly_chart(fig, use_container_width=True)

            st.warning("""**Note**: To improve the results of a model, practionners often conduct *hyperparameter tuning*. 
                        It consists of trying different combination of the model's parameters to maximise the accuracy score. 
                        Hyperparameter tuning wasn't conduct here in order to insure the app doesn't lag.""")
            
        with tab2:

            df_feature_imp = pd.DataFrame({"variable":feature_names, "importance":feature_imp})
            df_feature_imp = df_feature_imp.groupby("variable").mean().reset_index()
            df_feature_imp["importance"] = df_feature_imp["importance"].round(2)
            df_feature_imp.sort_values(by=["importance"], ascending=False, inplace=True)
            
            fig = px.bar(df_feature_imp, x="importance", y="variable", color="importance")
            st.plotly_chart(fig, use_container_width=True)

    else: # only show results for knn
        st.markdown("#### Results")

        st.markdown("""The K-nearest-neighbor algorithm doesn't have a built-in solution to compute model explainability with `scikit-learn`.
                    You can use other python packages such as `SHAP` to compute explainability, which we didn't use  here since they usually take a long time to output results.""")
        
        if select_data == "Diabetes 👩‍⚕️":
                st.error("""**Important**: Note that Diabetes dataset only contains information on 768 patients. 500 patients don't have diabetes and 268 do have the disease. 
                            This small number of patient data explains why the model's performance isn't optimal. 
                            Additional data collection should be conducted to improve results, as well as hyperparameter tuning (see explanation after graph).""")
                
        score_df = pd.DataFrame({"label":labels, "accuracy":np.round(score*100)})
        fig = px.bar(score_df, x="label", y="accuracy", color="label", title="Accuracy results", text_auto=True)
        st.plotly_chart(fig, use_container_width=True)

        st.warning("""**Note**: To improve the results of a model, practionners often conduct *hyperparameter tuning*. 
                        It consists of trying different combination of the model's parameters to maximise the accuracy score. 
                        Hyperparameter tuning wasn't conduct here in order to insure the app doesn't lag.""")