# MetricFrame: Beyond Binary Classification#

This notebook contains examples of using MetricFrame for tasks which go beyond simple binary classification.

import functools

import numpy as np
import sklearn.metrics as skm

from fairlearn.metrics import MetricFrame


## Multiclass & Nonscalar Results#

Suppose we have a multiclass problem, with labels $$\in {0, 1, 2}$$, and that we wish to generate confusion matrices for each subgroup identified by the sensitive feature $$\in { a, b, c, d}$$. This is supported readily by MetricFrame, which does not require the result of a metric to be a scalar.

First, let us generate some random input data:

rng = np.random.default_rng(seed=96132)

n_rows = 1000
n_classes = 3
n_sensitive_features = 4

y_true = rng.integers(n_classes, size=n_rows)
y_pred = rng.integers(n_classes, size=n_rows)

temp = rng.integers(n_sensitive_features, size=n_rows)
s_f = [chr(ord("a") + x) for x in temp]


To use confusion_matrix(), we need to prebind the labels argument, since it is possible that some of the subgroups will not contain all of the possible labels

conf_mat = functools.partial(skm.confusion_matrix, labels=np.unique(y_true))


With this now available, we can create our MetricFrame:

mf = MetricFrame(
metrics={"conf_mat": conf_mat}, y_true=y_true, y_pred=y_pred, sensitive_features=s_f
)


From this, we can view the overall confusion matrix:

mf.overall

conf_mat    [[101, 98, 126], [113, 97, 93], [113, 134, 125]]
dtype: object


And also the confusion matrices for each subgroup:

mf.by_group

conf_mat
sensitive_feature_0
a [[28, 27, 24], [32, 17, 26], [26, 32, 36]]
b [[27, 23, 25], [26, 32, 30], [36, 34, 28]]
c [[20, 26, 34], [27, 29, 19], [23, 33, 30]]
d [[26, 22, 43], [28, 19, 18], [28, 35, 31]]

Obviously, the other methods such as group_min() will not work, since operations such as ‘less than’ are not well defined for matrices.

Metric functions with different return types can also be mixed in a single MetricFrame. For example:

recall = functools.partial(skm.recall_score, average="macro")

mf2 = MetricFrame(
metrics={"conf_mat": conf_mat, "recall": recall},
y_true=y_true,
y_pred=y_pred,
sensitive_features=s_f,
)

print("Overall values")
print(mf2.overall)
print("Values by group")
print(mf2.by_group)

Overall values
conf_mat    [[101, 98, 126], [113, 97, 93], [113, 134, 125]]
recall                                              0.322308
dtype: object
Values by group
conf_mat    recall
sensitive_feature_0
a                    [[28, 27, 24], [32, 17, 26], [26, 32, 36]]  0.321359
b                    [[27, 23, 25], [26, 32, 30], [36, 34, 28]]  0.336450
c                    [[20, 26, 34], [27, 29, 19], [23, 33, 30]]  0.328501
d                    [[26, 22, 43], [28, 19, 18], [28, 35, 31]]  0.302603


## Non-scalar Inputs#

MetricFrame does not require its inputs to be scalars either. To demonstrate this, we will use an image recognition example (kindly supplied by Ferdane Bekmezci, Hamid Vaezi Joze and Samira Pouyanfar).

Image recognition algorithms frequently construct a bounding box around regions where they have found their target features. For example, if an algorithm detects a face in an image, it will place a bounding box around it. These bounding boxes constitute y_pred for MetricFrame. The y_true values then come from bounding boxes marked by human labellers.

Bounding boxes are often compared using the ‘iou’ metric. This computes the intersection and the union of the two bounding boxes, and returns the ratio of their areas. If the bounding boxes are identical, then the metric will be 1; if disjoint then it will be 0. A function to do this is:

def bounding_box_iou(box_A_input, box_B_input):
# The inputs are array-likes in the form
# [x_0, y_0, delta_x,delta_y]
# where the deltas are positive

box_A = np.array(box_A_input)
box_B = np.array(box_B_input)

if box_A < 0:
if box_A < 0:
raise ValueError("Bad delta y for box_A")
if box_B < 0:
raise ValueError("Bad delta x for box_B")
if box_B < 0:
raise ValueError("Bad delta y for box_B")

# Convert deltas to co-ordinates
box_A[2:4] = box_A[0:2] + box_A[2:4]
box_B[2:4] = box_B[0:2] + box_B[2:4]

# Determine the (x, y)-coordinates of the intersection rectangle
x_A = max(box_A, box_B)
y_A = max(box_A, box_B)
x_B = min(box_A, box_B)
y_B = min(box_A, box_B)

if (x_B < x_A) or (y_B < y_A):
return 0

# Compute the area of intersection rectangle
interArea = (x_B - x_A) * (y_B - y_A)

# Compute the area of both the prediction and ground-truth
# rectangles
box_A_area = (box_A - box_A) * (box_A - box_A)
box_B_area = (box_B - box_B) * (box_B - box_B)

# Compute the intersection over union by taking the intersection
# area and dividing it by the sum of prediction + ground-truth
# areas - the intersection area
iou = interArea / float(box_A_area + box_B_area - interArea)

return iou


This is a metric for two bounding boxes, but for MetricFrame we need to compare two lists of bounding boxes. For the sake of simplicity, we will return the mean value of ‘iou’ for the two lists, but this is by no means the only choice:

def mean_iou(true_boxes, predicted_boxes):
if len(true_boxes) != len(predicted_boxes):
raise ValueError("Array size mismatch")

all_iou = [
bounding_box_iou(y_true, y_pred)
for y_true, y_pred in zip(true_boxes, predicted_boxes)
]

return np.mean(all_iou)


We need to generate some input data, so first create a function to generate a single random bounding box:

def generate_bounding_box(max_coord, max_delta, rng):
corner = max_coord * rng.random(size=2)
delta = max_delta * rng.random(size=2)

return np.concatenate((corner, delta))


Now use this to create sample y_true and y_pred arrays of bounding boxes:

def many_bounding_boxes(n_rows, max_coord, max_delta, rng):
return [generate_bounding_box(max_coord, max_delta, rng) for _ in range(n_rows)]

true_bounding_boxes = many_bounding_boxes(n_rows, 5, 10, rng)
pred_bounding_boxes = many_bounding_boxes(n_rows, 5, 10, rng)


Finally, we can use these in a MetricFrame:

mf_bb = MetricFrame(
metrics={"mean_iou": mean_iou},
y_true=true_bounding_boxes,
y_pred=pred_bounding_boxes,
sensitive_features=s_f,
)

print("Overall metric")
print(mf_bb.overall)
print("Metrics by group")
print(mf_bb.by_group)

Overall metric
mean_iou    0.121063
dtype: float64
Metrics by group
mean_iou
sensitive_feature_0
a                    0.113238
b                    0.129801
c                    0.118929
d                    0.121758


The individual entries in the y_true and y_pred arrays can be arbitrarily complex. It is the metric functions which give meaning to them. Similarly, MetricFrame does not impose restrictions on the return type. One can envisage an image recognition task where there are multiple detectable objects in each picture, and the image recognition algorithm produces multiple bounding boxes (not necessarily in a 1-to-1 mapping either). The output of such a scenario might well be a matrix of some description. Another case where both the input data and the metrics will be complex is natural language processing, where each row of the input could be an entire sentence, possibly with complex word embeddings included.

## Conclusion#

This notebook has given some taste of the flexibility of MetricFrame when it comes to inputs, outputs and metric functions. The input arrays can have elements of arbitrary types, and the return values from the metric functions can also be of any type (although methods such as group_min() may not work).

Total running time of the script: (0 minutes 0.090 seconds)

Gallery generated by Sphinx-Gallery