The dataset is acquired from kaggle which entitled COVID vaccination vs. mortality. The CSV file contains vaccination rate data and its datetime for each respective countries. In addition to that, the columns provided in this CSV are as follows:
- country
- provided country name
- iso_code
- iso_code for each country
- date
- date that this data belong
- total_vaccinations
- number of all doses of COVID vaccine usage in that country
- people_vaccinated
- number of people who got at least one shot of COVID vaccine
- peoplefullyvaccinated
- number of people who got full vaccine shots
- New_deaths
- number of daily new deaths
- population
- 2021 country population
- ratio
- % of vaccinations in that country at that date = people_vaccinated/population * 100.
The main concern when creating our Machine Learning model are country and ratio column.
Based on current dataset, the vaccination rate from 2021-01-28 to 2021-11-20 in Indonesia is illustrated as follows:
There are various options for time-series model creation. Here, we tried to implement LSTM.
LSTM is the new fix from traditional RNN whereas it allows the model to keep the important values and remove those which aren't; in long period of time by default.
As the model's complexity grows, the possibility overfitting occurence getting higher as well. Even though the architecture is suited specifically for the dataset, using correct loss function, and suitable metrics, there is a chance of data overfitting. Therefore, in addition to LSTM we also use Dropout Layer to prevent overfitting while data is being trained. Simply, dropout layer will acts as hidden layer which will be enabled-disabled while training the data.
In this project, the dropout value is 0.5, the analogy is illustrated as follows:
For thorough model creation, see more here
Created model will be saved into h5 format, and it will be processed further in predict_vaccination.py
Overall, the architecture of this model is LSTM layer as the input layer, then it goes to dropout layer with 0.5 value to increase output variance. Then Dense layer with 1 perceptron will be the its output layer.
The prediction of vaccination rate in Indonesia
The training and validation MAE is illustrated as follows:
This metric is used to define model faults or variance of bias of trained model towards data that will be tested
See other repos :
