MULTILABEL VACCINE PROBAILITY PREDICTION

LOADING DATASET ( TRAINING AND TESTING CSV )

• using pandas read_csv()

EDA AND HANDLING MISSING VALUES

• using df.isnull().sum() lists out no.of null values in each feature/column

• out of 36 features given (30 features has missing values)

• 36 feature also consist of different data types (int/float/object(string))

• In most of the binary features NULL value is replace by MODE since either of 1 class is dominant

• Remaining Categorical Values are encoded using LabelEncoder from Scikit Learn

• Each features is represented using sns.countplot() to visualize

• Heatmap is plotted at the end to view correlation b/w different features

  

MODEL

• Before model selection test_train_split() to divide testing and training data with shuffle=True and stratified=True which improves random shuffling in the dataset

• Total 5 models have been used

  • i)Logistic Regression
  • ii)Bernoulli Naive bayes
  • iii)SVM - RBF Kernel function
  • iv)Random Forest
  • v)XGBoost

• models are trained with only one label xyz_vaccine/h1n1_vaccine

• since the model has high imbalance for xyz_vaccine/h1n1_vaccine compared to seasonal_vaccine as shown below

  

  • model which performed good on this label can be used for prediction of another label ( seasonal_vaccine)
  • Generic code for any model implement using a scikit pipeline

      model_1 = make_pipeline(
      LogisticRegression(solver='sag') )
      model_1.fit(X_train_xyz, y_train_xyz)
      plot_roc_curve(model_1, X_test_xyz, y_test_xyz, name='logistic regression - xyz vaccine')
      plt.show()
  

EVALUATION

• By plotting roc_auc curve and comparing relative roc_auc_scores " Tree based model " outperformed in classification task than remaining models .

• Both Random forest and XGBoost gave roc_auc_score of "0.83" and "0.82" respectively.

  

CONCLUSION

• Tree based models are best for this classification task , eventhough we have attained roc_score of 0.83 this can be improved a lot by detail Data preprocessing process using KNNImputer and Iterative Imputer instead of simple mode replacement (which may be computationally intensive but gives a better result)