MLFundamentals  Evaluation Metrics
Table of Contents
Introduction
Building a model to predict continious values (e.g. the temperature) or membership of a class, is only useful for real world problems, when we can measure how good (or bad) our model predicts. Therefore in this notebook, you will get to know (and implement) the most basic concepts of evaluating machinelearning models.
Requirements
Knowledge
You should have a basic knowledge of:
 numpy
Suitable sources for acquiring this knowledge are:
Python Modules
By deep.TEACHING convention, all python modules needed to run the notebook are loaded centrally at the beginning.
# External Modules
import numpy as np
import hashlib
import matplotlib.pyplot as plt
def round_and_hash(value, precision=4, dtype=np.float32):
"""
Function to round and hash a scalar or numpy array of scalars.
Used to compare results with true solutions without spoiling the solution.
"""
rounded = np.array([value], dtype=dtype).round(decimals=precision)
hashed = hashlib.md5(rounded).hexdigest()
return hashed
Exercises
Confusion Matrix
A confusion matrix is often used in binary classification tasks where we only have 2 classes (positive, negative), but it can also be constructed when we have more classes. The green elemtents mark the correct classifications. In some cases the classes can be more similar to one another (e.g. C1 might less different to C2 than to C3), which here is indicated by the intensity of the red color.
The variable y_pred
holds the labels for 10 predicted examples. The variable y_true
contains the true labels. A 0
means positive, 1
means negative. If the first element in y_pred
and the first element in y_true
is both 0
, it is a true positive and so on...
Task:
Implement the function to calculate the confusion matrix for given data. You function should return a 2Dmatrix:
 confmatrix[0,0] should contain the number of _true positives
 confmatrix[1,1] should contain the number of _true negatives
 confmatrix[1,0] should contain the number of _false negatives
 confmatrix[0,1] should contain the number of _false positives
dataset_size = 20
### use this for random data
#y_pred = np.random.random_integers(low=0, high=1, size=dataset_size)
#y_true = np.random.random_integers(low=0, high=1, size=dataset_size)
### use this for fixed data (can be automatically evaluated with assertions)
y_pred = np.array([+0, +0, 1, 1, +0, +0, 1, +0, 1, +0, 1, 1, +0, +0, +0, 1, +0, 1, 1, 1])
y_true = np.array([1, +0, +0, 1, 1, 1, 1, 1, 1, +0, +0, 1, 1, 1, +0, +0, 1, 1, 1, +0])
print(y_pred)
print(y_true)
def get_confusion_matrix_2_classes(y_pred, y_true):
conf = np.ndarray([2,2], dtype=np.int32)
### Your implementation here
return conf
conf_matrix = get_confusion_matrix_2_classes(y_pred, y_true)
print(conf_matrix)
print(np.array([['tp', 'fp'], ['fn', 'tn']]))
assert round_and_hash(conf_matrix) == '52ca17a7de673a7e78903f6a8ea91a0c'
Sidenote
A confusion matrix contains nearly all relevant information about the predictions our model has made. However comparing two or more confusion matrices of different models directly is hardly practical. Especially for human readability / interpretability it is necessary to withdraw unnecessary information. The rest of this notebook is about metrics, which are directly comparable
Accuracy
Probably the most intuitive metric is the accuracy. The accuracy specifies what percentage of our predictions are correct.
For two classes:
$ accuracy = \frac{tp + tn}{tp + tn + fp + fn} $
And in the general case with $ n $ classes:
$ accuracy = \frac{\sum_{i=1}^n h_{ii}}{\sum_{i=1}^n\sum_{j=1}^n h_{ij}} $
Task:
Implement the function to calculate the accuracy for the two class case, based on your confusion matrix.
Optional:
Implement your function to be able to process confusion matrices with $ n $ classes, with $ n \ge 2 $
def calc_accuracy(conf_matrix):
raise NotImplementedError()
def calc_accuracy(conf_matrix):
correct = 0
for i in range(conf_matrix.shape[0]):
correct += conf_matrix[i,i]
return correct / np.sum(conf_matrix)
Sidenote
Although the accuracy is very intuitive it might not be suitable for all classification tasks. Problems are:
 Inbalanced classes. Class A has 1.000 times more examples than class B. A model which just classifies everything as Class A not even looking at any features reaches 99.9% accuracy but is hardly even callable a machinelearning model.
 When in addition the missclasification of a true class B example as class A has sever consequences. Let class A be fish in the water and class B naval mines. When you are in a submarine you rather want your naval mine detector to have 95% accuracy but correctly predicting true mines.
accuracy = calc_accuracy(conf_matrix=conf_matrix)
print(accuracy)
assert round_and_hash(accuracy) == 'eaff9b69a66af6d38e881cfcce709153'
Recall, Precision, Specificity and FallOut
The following metrics set a specific field of a binary class confusion matrix in relation to another field. They can be hard to memorize, but the picture might help with that.
Acronyms:

Recall $ \frac{TP}{TP + FN} $ :
 sensivity, hit rate, true positive rate (TPR)

Precision $ \frac{TP}{TP +FP} $ :
 positive predictive velue (PPV)

Specificity $ \frac{TN}{TN + FP} $ :
 selectivity, true negative rate (TNR)

Fallout $ \frac{FP}{FP +TN} $ :
 false positive rate (FPR)
The other combinationes negative predictive value (NPV) $ \frac{TN}{TN + FN} $ , false negative rate (FNR) $ \frac{FN}{FN +TP} $ , false discovery rate (FDR) $ \frac{FP}{FP + TP} $ and false omission rate (FOR) $ \frac{FN}{FN + TN} $ also exist, but the four mentioned in the heading and especially recall and precision are the ones mostly used.
Balanced Accuracy
Balanced accuracy should be favoured, when classes are inbalanced. It is calculated with:
$ accuracy_{balanced} = \frac{TPR + TNR}{2} $
Task:
Consider the confusion matrix with fish_and_mines_pediction
, which classifies all objects as fish (negative) and does not detect the mine (positive) as a mine (false negative).
Implement the method to calculate $ accuracy_{balanced} $ .
 What is the value for the accuracy?
 What is the value for the balanced accuracy?
fish_and_mines_prediction = np.array([[1,0],[1,999]])
print(fish_and_mines_prediction)
def calc_balanced_accuracy(conf_matrix):
raise NotImplementedError()
acc = calc_accuracy(fish_and_mines_prediction)
print('accuracy: ', acc)
balanced_acc = calc_balanced_accuracy(fish_and_mines_prediction)
print('balanced accuracy: ', balanced_acc)
assert round_and_hash(balanced_acc) == '028aa23bdcb0575befa15321df88425e'
F1Score
Another well known measure is the F1Score (also FScore / FMeasure). It is defined as:
$ F_1 = 2 \cdot \frac{precision \cdot recall}{precision + recall} $
Task:
 Implement the method to calculate the F1Score
 What is the F1Score for
fish_and_mines_prediction
for class mine? 
Wha tis the F1Score if we consider fish as positive class?

For this task also implement the function to reverse the positives and negatives of a confusion matrix
 Do not reverse the matrix inplace. Create a new one and return it.

def calc_f1_score(conf_matrix):
raise NotImplementedError()
def reverse_pos_and_neg_of_conf_matrix(conf_matrix):
raise NotImplementedError()
def calc_f1_score(conf_matrix):
cm = conf_matrix
tpr = cm[0,0] / (cm[0,0] + cm[1,0])
ppv = cm[0,0] / (cm[0,0] + cm[0,1])
return 2. * ((tpr*ppv) / (tpr+ppv))
def reverse_pos_and_neg_of_conf_matrix(conf_matrix):
cm = conf_matrix
cm_reverse = np.ndarray((2,2))
cm_reverse[0,0] = cm[1,1]
cm_reverse[1,1] = cm[0,0]
cm_reverse[0,1] = cm[1,0]
cm_reverse[1,0] = cm[0,1]
return cm_reverse
f1score_mine = calc_f1_score(fish_and_mines_prediction)
print(f1score_mine)
assert round_and_hash(f1score_mine) == '6602c18aac262cc65614adb27ed43b2d'
fish_and_mines_reverse = reverse_pos_and_neg_of_conf_matrix(fish_and_mines_prediction)
f1score_fish = calc_f1_score(fish_and_mines_reverse)
print(f1score_fish)
assert round_and_hash(f1score_fish) == 'd4d4c5275731712b629968e3e8d1951e'
ROC
In most models, there is one or more parameters, which allows you to adjust your model to predict more true positives, but in return also predicts more false positives. For our example with the fish and mines: Adjusting a parameter to correctly find true mines, will most likely also have the consequence, that some fish will also be predicted as mines.
The ReceiverOperatingCharacteristicCurve is a method to visualize this. To do so, you adjust your parameter(s) to, let us say, 10 different values and predict your data. For all 10 experiments you calculate the FPR and the TPR and plot the results in a 2DDiagram. If your resulting curve is a straigth line from the bottom left to the top right, then your model just predicts totally random, which is the worst case. The better your model is, the higher the slope at the beginning.
Picture inspired by [SHA15]
AUC
To summarize the ROC in a single number (in order to compare model), you calculate the AreaUnderTheCurve (AUC), which will be a value between 0.0 and 1.0. Though 0.5 is the worst (random predictions) model. A value less then 0.5 most likely means, you just unintendedly switched some numbers.
Task:
Implement the method to calculate the ROC value for given 4 matrices. The resulting plot should look simialr to the picture above.
def calc_roc_value(conf_matrix):
raise NotImplementedError()
return (tpr, fpr)
cm1 = np.array([[63,28],[37,72]])
cm2 = np.array([[73,43],[27,57]])
cm1 = np.array([[1,1],[99,99]])
cm2 = np.array([[60,40],[40,60]])
cm3 = np.array([[99,99],[1,1]])
cm1_roc = calc_roc_value(cm1)
cm2_roc = calc_roc_value(cm2)
cm3_roc = calc_roc_value(cm3)
tprs = [cm1_roc[0], cm2_roc[0], cm3_roc[0]]
fprs = [cm1_roc[1], cm2_roc[1], cm3_roc[1]]
print(tprs)
print(fprs)
plt.plot(np.linspace(0,1),np.linspace(0,1), label='random guess')
plt.plot(fprs, tprs, label='ROC')
plt.legend()
tpr, fpr = calc_roc_value(fish_and_mines_prediction)
print(tpr, fpr)
assert round_and_hash(tpr) == '38fc9331271e16f0e5586a0fc993be00'
assert round_and_hash(fpr) == 'f1d3ff8443297732862df21dc4e57262'
Kappa Score
The kappa score takes into consideration that some correct predictions were made by 'accident':
\begin{equation} \kappa = \frac{p0  p_e}{1  p_e}, \end{equation} with $ p_0 $ being the accuracy and $ p_e $ the proportion of 'accidentially'_ correct classified examples.
For the binary classification task $ p_e $ is calculated with:
\begin{equation} p_e = \frac{(TP + FN) \cdot (TP + FP)}{b^2} + \frac{(FN + TN) \cdot (FP + TN)}{b^2} \end{equation} with $ b $ the total number of examples.
And in general for (n) different classes:
\begin{equation} pe = \frac{1}{b^2} \cdot \sum{i=1}^{n} h{i+} \cdot h{+i} \end{equation}
with the sum of row $ i SINGLE``SINGLE h_{i+} $ and the sum of column $ i SINGLE``SINGLE h_{+i} $
Weighted Kappa Score
If some misclassificaitons are worse then others (C1 classified as C3 is worse than classified as C2), it is possible to take weights into calculation. In this case we assign weights $ w_{11} $ to $ w_{nn} $ to the confusion matrix. For the weighted kappa score we then have:
\begin{equation} \kappaw = 1  \frac{\sum_i^n \sum_j^n w{ij} \cdot h{ij}}{\sum_i^n \sum_j^n w{ij} \cdot \frac{h{i+} \cdot h{+j}}{b}} \end{equation}
Summary and Outlook
[TODO]
Literature
Licenses
Notebook License (CCBYSA 4.0)
The following license applies to the complete notebook, including code cells. It does however not apply to any referenced external media (e.g., images).
Exercise: Evaluation Metrics
by Klaus Strohmenger
is licensed under a Creative Commons AttributionShareAlike 4.0 International License.
Based on a work at https://gitlab.com/deep.TEACHING.
Code License (MIT)
The following license only applies to code cells of the notebook.
Copyright 2019 Klaus Strohmenger
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.