Sample Weight

Sample Weight assigns weight to input sample. Weight may be specified by input param class_weight or sample_weight_name. Output data instances will each have a weight value, which will be used for training. While weighted instances may be used for prediction(SampleWeight component will assign weights to instances if prediction pipeline includes this component), Evaluation currently does not take weights into account when calculating metrics.

If result weighted instances include negative weight, a warning message will be given.

Please note that when weight is not None, only weight_diff convergence check method may be used for training GLM.

If both `class_weight` and `sample_weight_name` are provided, values
from column of `sample_weight_name` will be used.

Param

sample_weight_param

Attributes

Classes

SampleWeightParam(class_weight=None, sample_weight_name=None, normalize=False, need_run=True)

Bases: BaseParam

Define sample weight parameters

Parameters:

Name	Type	Description	Default
class_weight	str or dict, or None, default None	class weight dictionary or class weight computation mode, string value only accepts 'balanced'; If dict provided, key should be class(label), and weight will not be normalize, e.g.: {'0': 1, '1': 2} If both class_weight and sample_weight_name are None, return original input data.	None
sample_weight_name	str	name of column which specifies sample weight. feature name of sample weight; if both class_weight and sample_weight_name are None, return original input data	None
normalize	bool, default False	whether to normalize sample weight extracted from sample_weight_name column	False
need_run	bool, default True	whether to run this module or not	True

Source code in python/federatedml/param/sample_weight_param.py

def __init__(self, class_weight=None, sample_weight_name=None, normalize=False, need_run=True):
    self.class_weight = class_weight
    self.sample_weight_name = sample_weight_name
    self.normalize = normalize
    self.need_run = need_run

Attributes

class_weight = class_weight instance-attribute

sample_weight_name = sample_weight_name instance-attribute

normalize = normalize instance-attribute

need_run = need_run instance-attribute

Functions

check()

Source code in python/federatedml/param/sample_weight_param.py

def check(self):

    descr = "sample weight param's"

    if self.class_weight:
        if not isinstance(self.class_weight, str) and not isinstance(self.class_weight, dict):
            raise ValueError(f"{descr} class_weight must be str, dict, or None.")
        if isinstance(self.class_weight, str):
            self.class_weight = self.check_and_change_lower(self.class_weight,
                                                            [consts.BALANCED],
                                                            f"{descr} class_weight")
        if isinstance(self.class_weight, dict):
            for k, v in self.class_weight.items():
                if v < 0:
                    LOGGER.warning(f"Negative value {v} provided for class {k} as class_weight.")

    if self.sample_weight_name:
        self.check_string(self.sample_weight_name, f"{descr} sample_weight_name")

    self.check_boolean(self.need_run, f"{descr} need_run")

    self.check_boolean(self.normalize, f"{descr} normalize")

    return True

---

Last update: 2022-07-13