Class-imbalance voorkomen

Sommige data heeft erg scheve uitkomsten – zoals bij een gegevensset over een zeldzame ziekte. Als je willekeurig splitst, kun je pech hebben. Stel dat alle zeldzame observaties in de testset zitten en geen in de trainingsset. Dan gaat je hele trainingsproces de mist in!

Gelukkig biedt de functie initial_split() een oplossing. In deze oefening ga je deze zogeheten class-imbalances opsporen en aanpakken.

Er is al code aanwezig om een splitobject diabetes_split te maken met 75% training en 25% test.

Deze oefening maakt deel uit van de cursus

Machine Learning met boomgebaseerde modellen in R

Cursus bekijken

Praktische interactieve oefening

Probeer deze oefening eens door deze voorbeeldcode in te vullen.

# Preparation
set.seed(9888)
diabetes_split <- initial_split(diabetes, prop = 0.75)

# Proportion of 'yes' outcomes in the training data
counts_train <- table(training(___)$outcome)
prop_yes_train <- counts_train["___"] / sum(counts_train)

# Proportion of 'yes' outcomes in the test data
counts_test <- table(___)
prop_yes_test <- ___ / sum(___)

paste("Proportion of positive outcomes in training set:", round(prop_yes_train, 2))
paste("Proportion of positive outcomes in test set:", round(prop_yes_test, 2))

Code bewerken en uitvoeren

Deze oefening maakt deel uit van de cursus

Machine Learning met boomgebaseerde modellen in R

SkillTag.level.beginnerSkillTag.label

4.9+

Begin de cursus gratis

Ready to build a real machine learning pipeline? Complete step-by-step exercises to learn how to create decision trees, split your data, and predict which patients are most likely to suffer from diabetes. Last but not least, you’ll build performance measures to assess your models and judge your predictions.

Exercise 1: Welkom bij de cursus!Exercise 2: Waarom methoden op basis van bomen?Exercise 3: Specificeer die tree Exercise 4: Train dat model Exercise 5: Zo laat je je boom groeien Exercise 6: Train/test-split Exercise 7: Class-imbalance voorkomen

Huidige oefening

Exercise 8: Van nul tot held Exercise 9: Voorspellen en evalueren Exercise 10: Voorspellingen maken Exercise 11: Kraak de matrix Exercise 12: Voorspel je het goed?

Ready for some candy? Use a chocolate rating dataset to build regression trees and assess their performance using suitable error measures. You’ll overcome statistical insecurities of single train/test splits by applying sweet techniques like cross-validation and then dive even deeper by mastering the bias-variance tradeoff.

Exercise 1: Continuous outcomes Exercise 2: Train a regression tree Exercise 3: Predict new values Exercise 4: Inspect model output Exercise 5: Performance metrics for regression trees Exercise 6: In-sample performance Exercise 7: Out-of-sample performance Exercise 8: Bigger mistakes, bigger penalty Exercise 9: Cross-validation Exercise 10: Create the folds Exercise 11: Fit the folds Exercise 12: Evaluate the folds Exercise 13: Bias-variance tradeoff Exercise 14: Call things by their names Exercise 15: Adjust model complexity Exercise 16: In-sample and out-of-sample performance

Time to get serious with tuning your hyperparameters and interpreting receiver operating characteristic (ROC) curves. In this chapter, you’ll leverage the wisdom of the crowd with ensemble models like bagging or random forests and build ensembles that forecast which credit card customers are most likely to churn.

Exercise 1: Tuning hyperparameters Exercise 2: Generate a tuning grid Exercise 3: Tune along the grid Exercise 4: Pick the winner Exercise 5: More model measures Exercise 6: Calculate specificity Exercise 7: Draw the ROC curve Exercise 8: Area under the ROC curve Exercise 9: Bagged trees Exercise 10: Create bagged trees Exercise 11: In-sample ROC and AUC Exercise 12: Check for overfitting Exercise 13: Random forest Exercise 14: Bagged trees vs. random forest Exercise 15: Variable importance

Ready for the high society of tree-based models? Apply gradient boosting to create powerful ensembles that perform better than anything that you have seen or built. Learn about their fine-tuning and how to compare different models to pick a winner for production.

Exercise 1: Introduction to boosting Exercise 2: Bagging vs. boosting Exercise 3: Specify a boosted ensemble Exercise 4: Gradient boosting Exercise 5: Train a boosted ensemble Exercise 6: Evaluate the ensemble Exercise 7: Compare to a single classifier Exercise 8: Optimize the boosted ensemble Exercise 9: Tuning preparation Exercise 10: The actual tuning Exercise 11: Finalize the model Exercise 12: Model comparison Exercise 13: Compare AUC Exercise 14: Plot ROC curves Exercise 15: Wrap-up