As slow as it takes I am finally pick up a dying trend (or maybe not) : COVID-19 Case Prediction Model! Predictions are important especially when it comes to anticipating a crisis e.g. food supply shortage, disease breakout, earthquake/volcanic eruptions etc. Dataset is readily available.

Reduce computation power by reducing nodes in network from 64-32 to 8-4 nodes in hidden layer 1 and hidden layer 2 respectively, model performance is similar:

Model optimization : Reduce MAPE to 0.09 by reducing window size to 7.

Model optimization : Reduce MAPE to 0.11 by increase training epoch to 1000.

Model is able to achieve Mean Absolute Percentage Error (MAPE) of value 0.14: (MSE-Mean Squared Error, MAE-Mean Absolute Error)

RNN model is constructed using 2 LSTM layers with activation function='tanh', having 64 and 32 nodes respectively:

Hyperparameter optimizer,loss function,metrics are summarised below:

model.compile(optimizer='adam',loss='mse',metrics=['mse','mae','mape'])

tb=TensorBoard(log_dir=LOG_PATH)

hist=model.fit(x_train,y_train,batch_size=64,epochs=100,callbacks=tb,verbose=1)

Train and test dataset are available in separate files, having shape of (680,31) and (100,31) respectively, and have following features:

'date', 'cases_new', 'cases_import', 'cases_recovered', 'cases_active','cases_cluster', 'cases_unvax', 'cases_pvax', 'cases_fvax','cases_boost', 'cases_child', 'cases_adolescent', 'cases_adult','cases_elderly', 'cases_0_4', 'cases_5_11', 'cases_12_17','cases_18_29', 'cases_30_39', 'cases_40_49','cases_50_59','cases_60_69', 'cases_70_79', 'cases_80', 'cluster_import','cluster_religious', 'cluster_community', 'cluster_highRisk','cluster_education', 'cluster_detentionCentre', 'cluster_workplace', whereby 'cases_new' is the target feature.

Upon inspection 'cases_new' in train dataset contain non-numeric values as such:

To convert non-numeric values to null values the following code is appplied: df['cases_new']=pd.to_numeric(df['cases_new'],errors='coerce')

12 null values observed in 'cases_new' are represented as white stripes in figure below:

The following code is appplied to impute null values in time series data:

df['cases_new']=df['cases_new'].interpolate(method='linear')

Before	After

Notice the little gaps on line plot (left figure) dissapeared after interpolated (right figure).

1 null value is identified and imputed using pandas.DataFrame.interpolate() method.

Train and test datasets are preprocessed in a similar fashion.

To normalize target feature:

df_scaled=mms.fit_transform(np.expand_dims(df['cases_new'],axis=-1))`
df_test_scaled=mms.transform(df_test_ori)

Window size of 30 is selected to create train and test datasets, these datasets consist of rows of consecutive 30-values list:

x_train=[]
y_train=[]

for i in range(win_size,np.shape(df_scaled)[0]):
    x_train.append(df_scaled[i-win_size:i])
    y_train.append(df_scaled[i])
    
# Increase dimension to fit into RNN model
x_train=np.array(x_train)
y_train=np.array(y_train)

For test dataset:

# Concatenate scaled train and test dataset
concat_df=np.concatenate([df_scaled,df_test_scaled],axis=0)

# Create test dataset
concat_test=concat_df[-130:]

x_test=[]

for i in range(win_size,np.shape(concat_test)[0]):
    x_test.append(concat_test[i-win_size:i])

x_test=np.array(x_test)

From the prediction plot we could see the model tries to keep up with the trend of real cases with minor 'lagness', and tends to represent as 'moving averages' when real cases fluctuates. Proceeding with the spike of real cases the performance of model reduced even more and not able to capture the rapid and fluctuations of real cases. As such, the model still has room for improvement although achieving evaluation MAPE value of 0.14. Some suggestions to improve include:

• Increase number of epoch.
• Reduce window size.
• Implement Bidirectional() LSTM layer.
• Apply moving average in data preprocessing step to smoothen the fluctuations.
• Apply lrDecay on Adam optimizer.
• Modify model architecture as such to incorporate CNN pooling layers.