Unbalanced class in KNN weight - cost-sensitive

Question

Hi! Im using sci-kit learn for fit and test a KNN method, and the dataset is unbalanced/desbalanced classes and i need apply weight to class but i dont know the function to use in KNN, in other methods you can use weight=balance_weights(y_train) but in KNN i think you need a function for calculate the distances but i dont know exacly how to the method use this function and how to define correctly the weights for unbalanced problem.

Sorry for my english.

Accepted Answer

Hi, in scikit's KNN, you can define the weights hyper-parameter of KNeighborsClassifier. One popular weight method for unbalanced datasets is to weight points by their inverse distances to the query point. An example of this is seen below,

neigh = KNeighborsClassifier(n_neighbors=3,weights='distance')

There are several other weight methods for KNN that are explained here,

http://scikit-learn.org/dev/modules/generated/sklearn.neighbors.KNeighborsClassifier.html

Typically, the distances are calculated using the euclidean distance, however, you can use other distance functions by setting the 'p' hyper-parameter of KNeighborsClassifier.

Unbalanced class in KNN weight - cost-sensitive

Subscription:

Related questions