[MADLIB-998] Class weights for SVM

• Type: New Feature
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Component/s: Module: Support Vector Machines
• Assignee: Unassigned
• Reporter: Rahul Iyer
• Created: 23/May/16 17:12
• Updated: 09/Jun/16 20:53
• Resolved: 09/Jun/16 20:52

Description

Add a class weight parameter to add weights to specific dependent variable values. This is useful for data with unbalanced classes i.e. situations where 1 class has (far) fewer data points compared to other class(es).

The general format will be similar to that in scikit-learn, described below:

class_weight: Sets the weight for the positive and negative classes. If not given, all classes are set to have weight one.

If class_weight = balanced, values of y are automatically adjusted as inversely proportional to class frequencies in the input data i.e. the weights are set as n_samples / (n_classes * bincount ( y )).
Alternatively, class_weight can be a mapping, giving the weight for each class.
Eg. For dependent variable values 'a' and 'b', the class_weight can be

{a: 2, b: 3}

. This would lead to each 'a' tuple's y value multiplied by 2 and
each 'b' y value will be multiplied by 3.

For regression, the class weights are always one.

'class_weight' will be part of the optional 'params' argument.

Issue Links

• links to: GitHub Pull Request #43