webapp.py

import streamlit as st
import numpy as np
import pandas as pd
import altair as alt
import pickle
import plotly.express as px
from datetime import datetime, date
from PIL import Image

# read the Malaysia states 
df_map = pd.read_csv('map/map.csv')

st.set_page_config(layout="wide")


def main():

    # Sidebar
    st.sidebar.title("Malaysia COVID-19 Cases and Vaccination")
    st.sidebar.write('''
        This project aims to gain insight into Malaysia's COVID-19 pandemic using data mining techniques.

        ---
    ''')
    st.sidebar.subheader("🧭Navigation")
    choice = st.sidebar.radio("go to", ('COVID-19 Dataset', 'Exploratory Data Analysis', 'Clustering Analysis', 'Regression', 'Classification', 'Time-Series Regression'), index=0)

    if choice == 'COVID-19 Dataset':
        page_dataset()
    elif choice == 'Exploratory Data Analysis':
        page_eda()
    elif choice == 'Clustering Analysis':
        page_clustering()
    elif choice == 'Regression':
        page_regression()
    elif choice == 'Classification':
        page_classification()
    elif choice == 'Time-Series Regression':
        page_time_series_regression()

    st.sidebar.write('---')
    st.sidebar.subheader("Presented by")
    st.sidebar.write(''' 
        Bing Quan, Kuan Yung and Jing Guan
    ''')


def page_dataset():

    st.title('🦠COVID-19 Dataset')
    st.markdown('''
        Our data was obtained from [COVID-19 Open Data by the Minister of Health (MoH)](https://github.com/MoH-Malaysia/covid19-public) and [Vaccination Data by COVID-19 Immunisation Task Force (CITF)](https://github.com/CITF-Malaysia/citf-public)
        
        Note: Basic data cleaning and pre-processing was done in a set of separate Jupyter notebooks. 
    ''')
    st.markdown('''
        ### Categories
        The two dataset were combined and categorized into `Epidemic`, `Vaccination and Registration`, and `MySejahtera` according to their purposes.
    ''')
    
    col_1, col_2, col_3 = st.columns(3)
    
    with col_1:
        with st.expander('Epidemic', expanded=True):
            st.info('''
                #### Epidemic
                This category contains all files related to the general epidemic. For example the number of daily cases, tests and deaths in each state and also in Malaysia as a whole. Includes all COVID-19 clusters, hospitals, ICU and PKRC (Quarantine and Treatment Centre) data too.
                #
                Files that are included in this category: 

                * cases_malaysia.csv
                * cases_state.csv
                * tests_malaysia.csv
                * tests_state.csv
                * deaths_malaysia.csv
                * deaths_state.csv
                * cluster.csv
                * hospital.csv
                * icu.csv
                * pkrc.csv

                #
            ''') # bad formatting idea 
    with col_2:
        with st.expander('Vaccination and Registration', expanded=True):
            st.info('''
                #### Vaccination and Registration
                Data related to vaccination and registration. Records the number of people daily registration for vaccination and the vaccines given to Malaysians on a daily basis. Contains data recording the side effects of different vaccines too.
                #
                Files that are included in this category: 

                * vax_malaysia.csv
                * vax_state.csv
                * aef.csv
                * aefi_serious.csv
                * vaxreg_malaysia.csv
                * vaxreg_state.csv

                # 
                # 
                #
                # 
            ''') 
    with col_3:
        with st.expander('MySejahtera', expanded=True):
            st.info('''
                #### MySejahtera

                This category contains all data obtained from the MySejatera app. This includes the number of unique check-ins in each state and also in the whole Malaysia. It also records the number of unique check-in locations.
                ##
                ###
                ###
                Files that are included in this category: 

                * checkin_malaysia.csv
                * checkin_state.csv
                * checkin_malaysia_time.csv
                * trace_malaysia.csv   
                
                #
                #
                #
                #
                #
                ###
            ''')

    st.write('### Data Cleaning')
    st.success('''
        #### Dataset without any missing value
        ##### Epidemic
        * cases_state.csv
        * tests_malaysia.csv
        * tests_state.csv
        * deaths_malaysia.csv
        * deaths_state.csv
        * clusters.csv

        ##### Vaccination and Registration
        * vax_malaysia.csv
        * vax_state.csv
        * aef.csv
        * aefi_serious.csv
        * vaxreg_malaysia.csv
        * vaxreg_state.csv

        ##### MySejahtera
        * checkin_malaysia.csv
        * checkin_state.csv
        * checkin_malaysia_time.csv
    ''')

    st.error('''
        #### Dataset with missing value
        ##### Epidemic
        * cases_malaysia.csv
        ##### Vaccination and Registration
        None
        ##### MySejahtera
        * trace_malaysia.csv
    ''')
    st.info('''
        #### Handling outliers
        Outlier are not being removed but kept. This is due to the outlier of our data are not an error. 
        However, moving average is used for time-series analysis to reduce "noises" caused by outliers.
    ''')
    st.warning('''
        #### Dropping Missing Value
        ##### Epidemic
        * cases_malaysia.csv splitted into 2 dataset, one retains the missing values, the other one removes the missing values.
        ##### Vaccination and Registration
        None
        ##### MySejahtera
        * trace_malaysia.csv
    ''')



def page_eda():

    st.title('🔎Exploratory Data Analaysis')
    st.markdown('''
        We came up with a set of questions to help us understand the data more.
    ''')
    isExpand = st.checkbox('Expand all questions')
    

    with st.expander('EDA 1 - Analyse which group of population are more vulnerable to COVID cases in Malaysia.', expanded=isExpand):
        st.subheader('Which group of population are more vulnerable to COVID cases in Malaysia.')

        col1, col2 = st.columns((5,5))
        im = Image.open('images/EDA01_1.png')
        col1.image(im, caption='COVID Cases Group By Age')
        im = Image.open('images/EDA01_2.png')
        col2.image(im, caption='COVID Cases Group By Vaccination Status')
        im = Image.open('images/EDA01_3.png')
        col1.image(im, caption='COVID Cases Group By Cluster Type')

        st.write('Based on the observation, we can deduce that the population that are aged around 18-59, unvaccinated, and active in workfleids have the highest risk in getting COVID-19.')


    with st.expander('EDA 2 - Analyse how COVID cases vary across time dimensions at different granularity.', expanded=isExpand):
        st.subheader('How COVID cases vary across time dimensions at different granularity.')
        
        im = Image.open('images/EDA02_1.png')
        st.image(im, caption='COVID Cases Over Weekday')
        st.write('There is no clear indication showing that any particular day have more COVID cases. Weekends and weekdays are close to each other as well.')

        col1, col2 = st.columns((5,5))
        im = Image.open('images/EDA02_2.png')
        col1.image(im, caption='COVID Cases Over Week Number 2020')
        im = Image.open('images/EDA02_3.png')
        col2.image(im, caption='COVID Cases Over Week Number 2021')
        st.write('The cases tend to spike when it nears the end of the year (shown in both weeks and months). However, do not compare 2020 with 2021 because 2020 has a relatively smaller y range.')

        col1, col2 = st.columns((5,5))
        im = Image.open('images/EDA02_4.png')
        col1.image(im, caption='COVID Cases Over Month 2020')
        im = Image.open('images/EDA02_5.png')
        col2.image(im, caption='COVID Cases Over Month 2021')
        st.write('Same oberservation based on week number, the cases tend to spike when it nears 3rd or 4th quater of the year. But beware the the relative scale of 2020 is not the same as 2021')
        
    
    with st.expander('EDA 3 - What is the stationarity of the time-series dataset?', expanded=isExpand):
        st.subheader('What is the stationarity of the time-series dataset?')

        st.write('''In Econometrics, a stationary time series is one whose properties do not depend on the time at which the series is observed. Thus, time series with trends, or with seasonality,
         are not stationary — the trend and seasonality will affect the value of the time series at different times. In order to identify the stationarity of the time series, we use 
         [Dickey Fuller test](https://machinelearningmastery.com/time-series-data-stationary-python/#:~:text=series%20is%20stationary.-,Augmented%20Dickey-Fuller%20test,-Statistical%20tests%20make)
         ''')

        EDA3 = st.selectbox('Check Stationarity Of:', \
        ['COVID Cases', 'COVID Deaths', 'Cluster Tracing', 'MySejahtera Checkins'])

        if EDA3 == 'COVID Cases':
            im = Image.open('images/EDA03_1.png')
            st.image(im, caption='COVID Cases')
            st.markdown('''
            Reuslts of Dickey-Fuller Test for COVID Cases:

            | Test                           | Reuslts   |
            | ------------------------------ | ----------|
            | Test Statistic                 | -2.878612 |
            | p-value                        | 0.047892  |
            | #Lags Used                     | 19.000000 |
            | Number of Observations Used    | 600.000000|
            | Critical Value (1%)            | -3.441296 |
            | Critical Value (5%)            | -2.866369 |
            | Critical Value (10%)           | -2.569342 |

            Null hypothesis accepted, time series has non-stationarity
            ''')

        elif EDA3 == 'COVID Deaths':
            im = Image.open('images/EDA03_3.png')
            st.image(im, caption='COVID Deaths')
            st.markdown('''
            Reuslts of Dickey-Fuller Test for COVID Deaths:

            | Test                           | Reuslts   |
            | ------------------------------ | ----------|
            |Test Statistic                  |-3.412172  |
            |p-value                         | 0.010551  |
            |#Lags Used                      |19.000000  |
            |Number of Observations Used     |548.000000 |
            |Critical Value (1%)             |-3.442339  |
            |Critical Value (5%)             |-2.866829  |
            |Critical Value (10%)            |-2.569587  |

            Null hypothesis accepted, time series has non-stationarity
            ''')

        elif EDA3 == 'Cluster Tracing':
            im = Image.open('images/EDA03_4.png')
            st.image(im, caption='Covid Trace')
            st.markdown('''
            Reuslts of Dickey-Fuller Test for Covid Deaths:

            | Test                           | Reuslts   |
            | ------------------------------ | ----------|
            |Test Statistic                  |-2.089453  |
            |p-value                         | 0.248777  |
            |#Lags Used                      | 7.000000  |
            |Number of Observations Used     |160.000000 |
            |Critical Value (1%)             |-3.471896  |
            |Critical Value (5%)             |-2.879780  |
            |Critical Value (10%)            |-2.576495  |

            Null hypothesis accepted, time series has non-stationarity
            ''')
        elif EDA3 == 'MySejahtera Checkins':
            im = Image.open('images/EDA03_2.png')
            st.image(im, caption='MySejahtera Checkins')
            st.markdown('''
            Reuslts of Dickey-Fuller Test: for Covid Checkins

            | Test                           | Reuslts   |
            | ------------------------------ | ----------|
            |Test Statistic                  |-1.134774  |
            |p-value                         | 0.701007  |
            |#Lags Used                      |14.000000  |
            |Number of Observations Used     |294.000000 |
            |Critical Value (1%)             |-3.452790  |
            |Critical Value (5%)             |-2.871422  |
            |Critical Value (10%)            |-2.572035  |

            Null hypothesis accepted, time series has non-stationarity
            ''')


    with st.expander('EDA 4 - What are the vaccination and registration rates per state in Malaysia?', expanded=isExpand):
        st.subheader('What are the vaccination and registration rates per state in Malaysia?')
        col1, col2 = st.columns(2)
        with col1:
            im = Image.open('images/EDA04_1.png')
            st.image(im, caption='Vaccination percentage of each state')
        with col2:
            im = Image.open('images/EDA04_2.png')
            st.image(im, caption='Vaccination registration percentage of each state')

        st.markdown('''
            The figures above show the vaccination rate and vaccine registration rate per state. The figure on the left side is the vaccination rate and the right side is the vaccine registration rate for each state. According to the figures above, we could find some of the states have vaccination rates or vaccine registration rates that are larger than 100%, for example like Kuala Lumpur. For these situations, we consider foreigners' information may also be included in the dataset. 
            
            If we exclude those states over 100% of the rate, we found no states have achieved 80% of vaccination rate, Negeri Sembilan has the highest vaccination rate which is about 70%. But, for Sabah and Kelantan, their vaccination rate is still below or about 50% of the vaccination rate, which could be considered a low vaccination rate if compared with other states, and also the vaccine registration rate for Sabah and Kelantan are also low if compared with other states. 
            
            The government should pay attention to this situation to enhance their vaccination and registration rate to make sure each state could achieve the required vaccination rate to prevent the spread of the COVID virus.
        ''')


    with st.expander('EDA 5 - What are the types and total number of side effects for each type of vaccine?', expanded=isExpand): 
        st.subheader('What are the types and total number of side effects for each type of vaccine?')
        col1, col2 = st.columns([1, 2])
        with col1:
            st.markdown('''
                The figure on the left the total number of different side effects in dose 1 and dose 2 for each type of vaccine. According to the figure above, we could find the number of different types of side effects in dose one is significantly greater than the side effects in dose 2 for each type of vaccine above. Besides, we also found site pain, tiredness and headache are the most common side effects in each type of vaccine. To be aware, the dataset didn’t record any total amount of side effects for Cansino, so we can’t observe the information of side effects for Cansino.
            ''')
        with col2:
            im = Image.open('images/EDA05_2.png')
            st.image(im, caption='Barplot shows different types of side effects from each type of vaccine')

        st.write('')
        st.write('')
        st.write('')

        col3, col4 = st.columns([1, 2])
        with col3:
            st.markdown('''
                The next figure below also shows the number of serious side effects for each type of vaccine. No information regarding Cansino was recorded in the dataset. As we can see, actual facial paralysis is the largest number of serious side effects that exist in each type of vaccine. For Pfizer and Sinovac, suspected anaphylaxis is the second large number of serious side effects. 
            ''')
        with col4:
            im = Image.open('images/EDA05_3.png')
            st.image(im, caption='Barplot shows different types of serious side effects from each type of vaccine')


    with st.expander('EDA 6 - Which type of vaccine is given to more people?', expanded=isExpand):
        st.subheader('Which type of vaccine is given to more people?')
        col1, col2 = st.columns([2, 1])
        with col1:
            st.markdown('''
                The figure shows the ratio of each type of vaccine that is obtained by Malaysian residents.  As we can see, Pfizer is the major vaccine that is available in Malaysia, currently, a total of 51.43% of residents in Malaysia obtained Pfizer. In Addition, a total of 39.67% of residents obtained Sinovac, which is the second-largest amount of vaccine provided to Malaysian residents, followed by Astrazena and Cansino. Cansino is the smallest amount of vaccine obtained in Malaysia, currently, only 0.44% of Malaysia residents obtained Cansino.
            ''')
        with col2:
            im = Image.open('images/EDA06_1.png')
            st.image(im, caption='Pie Chart showing ratio of different types of vaccine given to Malaysians')


    with st.expander('EDA 7 - Which states are recovering? Which of the states shows a decrease in the number of COVID-19 cases?', expanded=isExpand):
        st.subheader('Which states are recovering? Which of the states shows a decrease in the number of COVID-19 cases?')
        st.markdown('''
            Since the number of vaccinated entities are increasing, there is a significant decrease in the number of COVID-19 cases in each state. By plotting the moving average of daily confirmed cases and the cumulative number of vaccinated persons against the dates, we can observe that there is a downward trend after the vaccination has reached a certain level in each of the states. The moving average of daily cases is obtained by averaging the number of daily cases in the past 7 days. 

            To calculate the daily cumulative number of vaccinated persons, we start by calculating the number of people who completed their vaccination dose. This includes adding the 

            * daily number of adults who has completed their 2nd dose of vaccine (either Pfizer, Sinovac or Astra)
            * daily number of adults who has taken Cansino (since Cansino only requires 1 dose)
            * daily number of children who complete their 2nd dose of Pfizer

            These numbers are accumulated to obtain the daily cumulative number of vaccinated persons.    
        ''')

        states_dict = {
            'Johor': 1,
            'Kedah': 2,
            'Kelantan': 3,
            'Melaka': 4,
            'Negeri Sembilan': 5,
            'Pahang': 6,
            'Perak': 7,
            'Perlis': 8,
            'Pulau Pinang': 9,
            'Sabah': 10,
            'Sarawak': 11,
            'Selangor': 12,
            'Terengganu': 13,
            'W.P. Kuala Lumpur': 14,
            'W.P. Labuan': 15,
            'W.P. Putrajaya': 16
        }

        states = st.multiselect(
            'Select states to plot the graph (select more to plot more)',
            list(states_dict.keys()),
            ['Selangor']
        )

        for state in states:
            idx = states_dict.get(state)
            im = Image.open('images/EDA07_{:02d}.png'.format(idx))
            st.image(im)

        st.markdown('''
            All of the states in Malaysia are gradually recovering. Most states show a decrease in the number of COVID-19 cases when the percentage of vaccinated persons reaches around 28% ~ 47%. Labuan and Negeri Sembilan started to recover at a surprisingly low vaccination rate of 12%. On the other hand, cases in Sarawak starts to decline at 64% while Kuala Lumpur at 54%. 

            The number of vaccinated people in some states has exceeded their state population. The two states that possess this phenomena are Kuala Lumpur and Putrajaya. This phenomenon might be due to foreigners receiving their vaccination shot in our country.
        ''')


    with st.expander('EDA 8 - When is the time of the day with most MySejahtera check-ins?', expanded=isExpand):
        st.subheader('When is the time of the day with most MySejahtera check-ins?')
        st.markdown('''
            checkin_malaysia_time.csv records the time distribution of daily check-ins on MySejahtera at the country level. The data records the number of check-ins nationwide in every 30 minutes interval.

            By summing up each column in the dataset, we can obtain the total number of check-ins at each 30 minutes between 2020-12-01 to 2021-10-05, spanning over 309 days.
        ''')
        im = Image.open('images/EDA08_1.png')
        st.image(im, caption='Total number of check-ins from MySejahtera data')
        st.write('From the plot above, we can roughly observe 3 peaks, which are around 8am, 12pm, and 6pm respectively. Presumably these are the preferred hours for Malaysians to buy their meals.')


    with st.expander('EDA 9 - What are the dates with the highest number of checkins? How does it correlate with the number of cases and deaths during the day?', expanded=isExpand):
        st.subheader('What are the dates with the highest number of checkins? How does it correlate with the number of cases and deaths during the day?')
        im = Image.open('images/EDA09_1.png')
        st.image(im, caption='Daily number of confirmed cases, deaths, and check-ins')

        col1, col2 = st.columns(2)
        with col1:
            st.markdown('''
                The date with the highest number of checkins between 2020-12-01 and 2021-10-02 is 2021-04-03 (the 3rd of April 2021). On that day, the number of cases and deaths are considerably low. 
                
                In order to visualize the correlation between the daily number of cases and deaths with the daily number of check-ins, we have plotted a heatmap among these 3 attributes. 
                
                The heatmap shows that there is almost no correlation between them.
            ''')
        with col2:
            im = Image.open('images/EDA09_2.png')
            st.image(im, caption='Correlation heatmap')


    with st.expander('EDA 10 - Rate of Serious Vaccine Side Effect  VS COVID Death Rate without obtaining vaccine, which one is more dangerous?', expanded=isExpand):
        st.subheader('Rate of Serious Vaccine Side Effect  VS COVID Death Rate without obtaining vaccine, which one is more dangerous?')
        st.markdown('''
            Some people refuse to obtain vaccines because they think vaccines are dangerous and high risk. In this question, we will measure the rate of serious side effects of vaccination and COVID death rate if without obtaining a vaccine to evaluate whether reducing it is a good choice or not. 
            
            We divide the total number of COVID deaths without obtaining a vaccine and divide with the total number of people without obtaining a vaccine then multiply with 100 to obtain the rate of COVID death without obtaining the vaccine. Besides, we sum out the number of serious side effect cases and divide the total number of people obtaining the vaccine then multiply it to 100 to obtain the rate of the serious side effect. We use both rates to multiply 1000 to obtain the number of death or serious side effects cases for every 1000 people. The figure above shows the results of the measures. As we can see, for every 1000 people, about 200 people without obtaining a vaccine will die because of COVID. But only about 2 people may get a serious vaccine side effect for every 1000 people. 
        ''')
        col1, col2 = st.columns(2)
        with col1:
            im = Image.open('images/EDA10_1.png')
            st.image(im, caption='Number of cases and serious side effect of vaccine amoung 1000 people')
        with col2:
            im = Image.open('images/EDA10_2.png')
            st.image(im, caption='Raio of serious side effect in ratio of number of people not obtain vaccine and dead because of COVID (among 1000 people)')


        st.markdown('''
            The result at the left is evidence to show it is more dangerous if you do not obtain any vaccine. The figure  at the right shows the ratio of serious vaccine side effect cases in the number of COVID deaths without obtaining any vaccine for every 1000 people. We can find a ratio of serious vaccine side effects not greater than 1%. So that, in a nutshell, obtaining a vaccine is a safer and smart decision to protect your life. 
        ''')


def page_clustering():

    st.title('🧩Clustering Analysis')
    st.write('## **How well does each state handle COVID-19 cases based on past COVID-19 cases and deaths records?**')
    st.markdown('''
        To evaluate how well each state handles COVID-19 cases, we would like to perform time-series clustering to divide each state into different clusters based on their daily cases and deaths to analyze or discover the pattern or characteristics of each state. Time series k mean from `tslearn` library will be used to perform time-series clustering. 
        
        Before we start clustering the states, the number of daily cases and deaths of each state have been normalized by their state’s population and 3 clusters have been chosen (refer to our code for more) for the time series k means.

        Select a parameter for clustering:
    ''')

    df_map.cluster_cases = df_map.cluster_cases.astype('string')
    df_map.cluster_deaths = df_map.cluster_deaths.astype('string')
    
    cluster = st.selectbox('Clustering based on:', ['Daily COVID-19 Cases', 'Daily Number of Deaths'])

    if cluster == 'Daily COVID-19 Cases':
        cluster_attr = 'cluster_cases'
    else:
        cluster_attr='cluster_deaths'
    
    fig = px.scatter_mapbox(df_map, lat="lat", lon="lon", hover_name='state', 
        hover_data=["population"], color=cluster_attr,
        center={
            'lat': 4.0,
            'lon': 108.25
        }, 
        zoom=4.8, height=600,
        mapbox_style="carto-darkmatter",
        title=f"Clustering Malaysia states based on {cluster}",
        category_orders={cluster_attr: ['0', '1', '2']}
    )

    st.plotly_chart(fig, use_container_width=True)

    col1, col2 = st.columns([2, 1])
    with col1:
        
        if cluster == 'Daily COVID-19 Cases':
            st.subheader('Clustering each states into 3 clusters according to daily cases')
            im = Image.open('images/CTR_1.png')
            st.image(im, width=650, caption='Dividing each states into different clusters according to cases')
        else:
            st.subheader('Clustering each states into 3 clusters according to daily deaths')
            im = Image.open('images/CTR_2.png')
            st.image(im, width=650, caption='Dividing each states into different clusters according to deaths')

    with col2:
        st.subheader('States and their clusters')
        st.table(df_map[['state', cluster_attr]])

    st.write('') 
    st.write('')
    st.write('')
    
    if cluster == 'Daily COVID-19 Cases':
        st.markdown('''
            The figure above shows the results of clustering according to the daily cases. There is 1 state cluster as red clusters. Which is Labuan. The trends of daily cases slightly increase between 2020 October to 2021 March and significantly increase when 2021 June. In 2021 July, the cases start decreasing.  

            Next, there are 4 states as green clusters. Which are Kuala Lumpur, Melaka, Selangor and Negeri Sembilan. The cases of these states are quite stable until 2021 Jun, the cases start increasing until 2021 July, then cases decrease. 

            Lastly, 11 states are red clusters. The trends of cases for these states are also similar to states as green clusters, but the number of cases increases a little bit higher than states in green clusters.

            According to the result of clustering, except Labuan, before June 2021, each of the state's daily cases could be considered stable, most of the states were able to control the daily cases well. But in  June 2021, the cases for each state kept increasing, especially Labuan, whose cases have significantly increased if compared with other states. We can consider that during that period, Labuan was out of control in daily cases. But fortunately, after only about 1 month, Labuan cases have been under control, the cases start decreasing and during that period, other states' daily cases still keep increasing until September then start decreasing. In conclusion, some investigation on Labuan in handling COVID-19 cases should be considered especially in Jun 2021 since that time performs a very significant increase of cases which can’t be observed in other states. Besides, states in green clusters and states in red clusters have similar patterns but states in green clusters cases are less than the states in red clusters, which may handle the daily cases better if compared with those states in red or red clusters.
        ''')
    else:
        st.markdown('''
            The figure above shows the states clustering based on daily deaths cases. As we can see there is one state as a red cluster. The deaths slightly increase from September 2020 to March 2021 and a significant increase in deaths cases in June 2021. Besides, there are four states as green clusters, which are Kuala Lumpur, Melaka, Selangor, and Negeri Sembilan again. The death cases in green clusters keep increasing more rapidly if compared with states in blue clusters from June to September 2021. Furthermore, states in blue clusters perform more stable trends of death cases if compared with another two clusters. 

            According to this figure, we can conclude that Labuan is not doing well in handling COVID cases. In both figures, it performed a very significant increase in June 2021, and from November to March 2021, it also slightly increased its death cases but other states didn’t have the same problems. For states in the blue cluster, although their daily cases are slightly higher than states in the green cluster, their death cases are much more stable if compared with states in the green cluster. So, states in the blue cluster perform better in handling COVID cases and problems.       
        ''')

    st.write('') 
    st.write('')
    st.write('')

    with st.expander('💡 Tips'):
        st.info('Select clustering based on a differnt parameter above to get another set of results.')

        
def page_regression():

    st.title('📈Regression Models')
    st.write('''
    ## By utilizing the previous COVID-19 records, is it possible to construct a model capable of predicting the number of cases for the upcoming day or week?

    In another words, how can we use previous week's data (data from past 7 days) to predict or forecast number of covid cases for next day or next week using 
    [supervised learning for time series forecasting](https://machinelearningmastery.com/time-series-forecasting-supervised-learning/).''')
    st.write('### Data Preprocessing')
    st.write(''' 
    1. The data is first being cleaned and merged data from `cases`, `tests`, `deaths`, and `vaccination`. The date stamps are 
    **2020-01-31** to **2021-10-02**, 611 rows in total. 
    2. The missing values are filled with 0 because deaths and vaccination only started at some point and hence they are not an error or outlier. 
    3. Then normalize all attributes and calculate their moving average using a sliding window calculation.
    4. The moving average of covid cases of next day is duplicated and shifted 1 row up. See example below:

    | Date      | cases   | cases_duplicated|
    |-----------|---------|-----------------|
    | 22/1/2021 | 10      | 10              |
    | 23/1/2021 | 20      | 20              |
    | 24/1/2021 | 30      | 30              |

    Then "cases_duplicated" will undergo `Shift(-1)`

    | Date      | cases   | cases_duplicated|
    |-----------|---------|-----------------|
    | 22/1/2021 | 10      | 20              |
    | 23/1/2021 | 20      | 30              |
    | 24/1/2021 | 30      | data from 25/1  |

    "cases_duplicated" become covid cases of next day.

    5. The next_day data (aka. y_train) is [sample-binned using the median](https://towardsdatascience.com/data-preprocessing-with-python-pandas-part-5-binning-c5bd5fd1b950) 
    for classification later.''')

    X_train = pickle.load(open('pickle_files/classification_X_train.pkl', 'rb'))
    y_train = pickle.load(open('pickle_files/classification_y_train.pkl', 'rb'))
    
    st.write('#### X train, dimensions')
    st.write(X_train.head())
    st.write('#### y train, to be predicticted')
    st.write('Binned moving average of next day')
    st.write(y_train.head())
    st.write('Another worth mentioning part is the y train is being binned using median sample bin.')
    
    col1, col2 = st.columns((5,5))
    im1 = Image.open('images/DM_class_reg_data_median_bin.png')
    col1.image(im1, width=400, caption='5 equally distributed median')
    im2 = Image.open('images/DM_class_reg_data_median_bin2.png')
    col2.image(im2, width=400, caption='Sample bin median')

    st.write('#### Feature Importance')
    col1, col2 = st.columns([8, 9])
    im = Image.open('images/DM_class_reg_heatmap.png')
    col1.image(im, caption='Heatmap for All Features')
    im = Image.open('images/DM_class_reg_SHAP1.png')
    col2.image(im, caption='SHAP values of each features')
    im = Image.open('images/DM_class_reg_SHAP2.png')
    col2.image(im, caption='SHAP values of each features 2')

    st.markdown('''
    ### Regression Models

    Regression models that will be used:\n
    1. Linear Regression
    2. Decision Tree Regressor
    3. Random Forest Regressor
    4. Support Vector Regressor

    Evaluation matrics that will be used:\n
    1. R Square
    2. Mean Absolute Error(MAE)
    3. Root Mean Square Error(RMSE)
    ''')

    st.write('''
    Example for Support Vector Regressor\n
    ```Python
    reg_sv = SVR(kernel='linear')\n
    reg_sv = reg_sv.fit(X_train, np.ravel(y_train))\n
    y_pred = reg_sv.predict(X_test)
    ```
    ''')

    st.markdown('''
    #### Results
    |                               | Linear Regression | Decision Tree | Random Forest | SVR       |
    | ----------------------------- | ----------------- | ------------- | ------------- | --------- |
    | R Square                      | 0.89642           | 0.98097       | 0.93570       | 0.88917   |
    | Mean Absolute Error (MAE)     | 0.40572           | 0.04065       | 0.03211       | 0.39836   |
    | Root Mean Square Error (RMSE) | 0.47043           | 0.20162       | 0.14114       | 0.48661   |
    ''')

    st.write('')
    st.write('')

    st.write('''
    #### Conclusion
    An interesting observation is that Random Forest and Decision Tree both perform better than SVR and Linear Regressor.
    We decude that one of the reason behind this is both RF and DT supports non linearity better.
    ''')
    
def page_classification():

    st.title('📊Classification')
    st.write('''
    ## By utilizing the previous COVID-19 records, is it possible to construct a model capable of classifying the number of cases for the upcoming day or week?

    This is the continuation of Regression using similar [dataset](#data-preprocessing). A gentle reminder, do visit 
    regression page before diving into this classification section.
    ''')

    st.write('''
    ### Classification Models

    Classification models that will be used:\n
    1. K-Nearest Neighbors Classifier
    2. Naive Bayes Classifier
    3. Decision Tree Classifier
    4. Random Forest Classifier
    5. Support Vector Classifier

    Performance evaluation metrics that will be used:\n
    1. Confusion Matrix
    2. Precision, Recall & F1-Score

    ''')

    st.write('### Model Accuracy')
    st.write('''
        |                        | KNN       | GaussianNB  | Decision Tree | Random Forest | SVC      |
        | ---------------------- | ----------| ------------| ------------- | ------------- | -------- |
        | Test Set Accuracy      | 0.94309   | 0.95122     | 0.95935       | 0.95935       | 0.78049  |
    ''')

    st.write('')
    st.write('')

    st.write('### Confusion Matrix')
    im = Image.open('images/DM_class_reg_confusionMatrix.png')
    st.image(im)
    im = Image.open('images/DM_class_reg_confusionMatrix2.png')
    st.image(im, caption='Confusion Matrix of each model')

    st.write('')
    st.write('')

    st.write('### Precision, Recall & F1-Score')
    im = Image.open('images/DM_class_reg_ROC1.png')
    st.image(im)
    im = Image.open('images/DM_class_reg_ROC2.png')
    st.image(im, caption='Classification Report of each model')

    st.markdown('''
    ### Observation
    1. Decision tree and random forest outperforms other classifiers and regressors. The main reason could be they handle non-linearity and colinearity more efficiently. 
    2. The poor performance of SVM could because SVM suited two-class prediction better. Hence, for supervised time-series machine learning, our experiment result shows that decision tree and random forest performs better. 
    ''')

    st.markdown('''
    ### Future Improvement
    1. Includes other features that are not captured in dataset, for instance, date where government execute MCO.
    2. The poor results from SVM could due to the imbalance of dataset. Suggest to use SMOTE or other sampling techniques to deal with imbalance issue.
    ''')


def page_time_series_regression():

    st.title('⏳Time-Series Regression')
    st.write('## **How can the Malaysian government predict the number of daily new cases accurately based on past data in order to deploy appropriate movement control measures?**')
    st.markdown('''
        With an accurate forecast of the daily new cases, the government would be able to control the vaccination rate to tackle the pandemic effectively. We have implemented a time-series regression to forecast the number of COVID-19 cases in Malaysia. Two LSTM-based RNN is built to aid us in this problem.
    ''')
    
    ### model 1
    st.subheader('LSTM-based RNN')
    col1, col2 = st.columns(2)
    with col1:
        st.markdown('''
            The first time-series regression model is a single variate LSTM-based RNN model. The number of daily cases is separated into a train and test set on 2021-07-01. The training set has a total of 523 days and the test set has a total of 97 days.

            A 5 layer LSTM-based RNN with a total of 71,051 trainable parameters has been implemented. Each layer apart from the dense layer has 50 units and a dropout rate of 0.2. After training the model, we used it to predict the test set. The model failed to predict the trend of the daily cases as shown below. The results have a root mean square error of 7179.562.

            | Evaluation                    | LSTM-Based RNN |
            | ----------------------------- | -------------- |
            | Root Mean Square Error (RMSE) | 7179.562       |
        ''')
    with col2:
        im = Image.open('images/LSTM_1.png')
        st.image(im, width=500, caption='Actual and Predicted results for COVID-19 Cases')

    ### model 2
    st.subheader('Multivariate LSTM-based RNN')
    st.markdown('''
        Our second time-series regression model is a multivariate LSTM-based regression. We have plenty of datasets that might have an impact on the number of daily COVID-19 cases. These datasets are cases_malaysia, test_malaysia, deaths_malaysia, checkin_malaysia, vax_malaysia, and vaxreg_malaysia which records the cases, tests, deaths, check-ins, vaccination and registration data in Malaysia.

        A simple feature selection was used to retrieve the attributes with higher correlation to the number of daily cases. We have generated heatmaps to visualize the correlationship of each attribute with `cases_new` (daily COVID cases). Before that, we will be merging a few datasets together. Since tests_malaysia, deaths_malaysia, and checkin_malaysia have a similar time range, we will merge them together along with cases_malaysia. On the other hand, vax_malaysia and vaxreg_malaysia have a relatively shorter time range. Hence, we will merge them with the `cases_new` column separately in another DataFrame.

        After observing the correlationship between `cases_new` and other attributes from various datasets (cases_malaysia, tests_malaysia, deaths_malaysia, checkin_malaysia, vax_malaysia and vaxreg_malaysia). We found that the features with higher correlation (>= 0.9 positively or negatively) are:
    ''')

    col1, col2 = st.columns([1,2])
    with col1:
        st.markdown('''
            `cases_recovered`,  
            `cases_active`,  
            `cases_pvax`,  
            `cases_child`,  
            `cases_adolescent`,  
            `cases_adult`,  
            `cases_elderly`,  
            `deaths_new`,  
            `deaths_new_dod`, and  
            `deaths_bid_dod`.
        ''')
    with col2:
        im = Image.open('images/LSTM_2.png')
        st.image(im, caption='Using Heatmap for Feature Selection')


    st.markdown('''
        The resulting dataset has 568 days of data. Similar to the above, we implemented a 5 layer LSTM-based RNN with a total of 73,051 trainable parameters. A lookback of 30 days is used to create our dataset, resulting in 538 days of data. We then get the last 50 days for prediction and the rest of the days for training and validation. The model came to an early stopping at 16 epoch after training and validation data were fed to it.

        After the training, we test our model by predicting the test data. The results show that the model is able to capture the overall downward trend of the number of daily cases in Malaysia but failed to capture the ups and downs in the details. The model has a root mean square error of 1143.778, which is an improvement from the previous model.

        | Evaluation                    | Multivariate LSTM-Based RNN |
        | ----------------------------- | --------------------------- |
        | Root Mean Square Error (RMSE) | 1143.778                    |


    ''')
    st.write('') 
    st.write('')
    st.write('')
    
    im = Image.open('images/LSTM_3.png')
    st.image(im, caption='Actual and Predicted results for COVID-19 Cases using Multivariate LSTM')
    
    st.write('---')
    st.subheader('Stationary Time-Series and Non-stationary Time-Series')
    st.markdown('''
        Time-series data can be categorized into stationary and non-stationary. Stationary time-series data tends to be more predictable as the mean and variance normally remains constant and does not possess any trends or seasonality. On the other hand, non-stationary data are considered harder to predict as it is the opposite of stationary data. The mean and variance tends to fluctuate, making it harder to predict its behavior. 

        The daily number of COVID-19 cases in Malaysia is a non-stationary time-series data, hence the prediction results are not quite accurate.    
    ''')

    col1, col2 = st.columns([8, 10])
    with col1:
        im = Image.open('images/LSTM_5_2.png')
        st.image(im, caption='Stationary Time-Series and Non-stationary Time-Series')
    with col2:
        st.write('#')
        st.write('#')
        im = Image.open('images/LSTM_6.png')
        st.image(im, caption='Overview of Daily Recorded COVID-19 Cases in Malaysia from 2020-01-25 to 2021-10-05')

    st.markdown('''
        Most time series regression models perform better forecasts with stationary time-series data. It is known that LSTM is capable of forecasting non-stationary data. However, stationary data will greatly boost its performance as it is much easier for the neural network to learn. 

        #
    ''')

    st.subheader('Future Improvements')

    st.markdown('''
        Since the forecasting result is not quite accurate, future extensions can be done to improve the prediction results. Future work includes:

        1. Transforming non-stationary data into stationary data.
        2. Choosing more robust models capable of predicting non-stationary data. ARIMA model is known for able to handle 2 types of non-stationarities (hidden trend and unit roots). The ARIMA model might yield a better result in this case. 
        3. Prepare a larger dataset. More insights might be able to observe from a larger dataset.

        #
        ---
        **References**:

        1. [Stationarity in Time Series Analysis Explained using Python](https://blog.quantinsti.com/stationarity/)
        2. [Time Series Analysis using ARIMA and LSTM(in Python and Keras)](https://medium.com/analytics-vidhya/time-series-analysis-using-arima-and-lstm-in-python-and-keras-part1-f987e11f9f8c)    
        3. [Automated Hybrid Time Series Forecasting: Design, Benchmarking, and Use Cases](https://www.researchgate.net/publication/348592737_Automated_Hybrid_Time_Series_Forecasting_Design_Benchmarking_and_Use_Cases)
        4. [How to Remove Non-Stationarity in Time Series Forecasting](https://towardsdatascience.com/how-to-remove-non-stationarity-in-time-series-forecasting-563c05c4bfc7)
    ''')

if __name__ == "__main__":
    main()