datacamp

here are my notes on the first 3 chapters of the datacamp courses i did today (a continuation of supervised learning with scikit-learn):

chapter 3: fine-tuning your model

• how good is your model?
  • confusion matrix for assessing classification performance
    • actual:, predicted:
    • :legitimate, :fraudulent
    • true negative (top left actual:legitimate & predicted:legitimate), false positive, false negative, true positive (bottom right - actual:fraudulent & predicted:fraudulent)
  • precision - metric which is the number of true positives divided by sum of all positive predictions
  • recall - number of true positive divided by sum of true positives and false negatives (also called sensitivity - correct predictions of most fraudulent transactions)
  • f1 score - mean of precision and recall
    • 2 * (precision * recall / precision + recall)
• logistic regression and the roc curve
  • logistic regression - used for classification problems, and outputting probabilities (produces a linear decision boundary (straight line if 2d))
  • default threshold of logistic regression = 0.5
  • roc curve (receiver operating characteristic curve) - helps visualise different threshold effecting true positive and false positive rates

chapter 4: preprocessing and pipelines

• preprocessing data
  • using scikit-learn has requirements
    • numeric data
    • no missing values
  • we may have to preprocess data using categorical values into numeric values to allow us to use scikit-learn with our data
• handling missing data
  • imputation - replace missing data with educated guesses