How can I achieve a scikit-learn confusion matrix without extra columns?

Question

I have a classifier that does binary classification. For training I take the data I trust. For testing, I take the data I trust and also some not so good data (real world data).

How do I get a confusion matrix without extra columns?

Here is my test code

import matplotlib.pyplot as plt
from sklearn.metrics import confusion_matrix, ConfusionMatrixDisplay

y_test = [0, 1, 1, 1, 2, 2, 3]
predictions = [0, 1, 1, 1, 0, 1, 0]

cm = confusion_matrix(y_test, predictions)

l = ["M", "F", "M?", "F?"]
disp = ConfusionMatrixDisplay(confusion_matrix=cm, display_labels=l)
disp.plot()
plt.show()
# I expect a 4x2 matrix here. How can I do it?

See? Unnecessary columns are here

Am I missing something?

Feel free to say that I am completely wrong.

Answer 1

Because you said "... my question ishow to make a 4x2 matrix." on comments I wrote this:

import seaborn as sns
import matplotlib.pyplot as plt
import numpy as np
from sklearn.metrics import confusion_matrix


y_test = [0, 1, 1, 1, 2, 2, 3]
predictions = [0, 1, 1, 1, 0, 1, 0]
unique_labels = sorted(set(y_test + predictions))

# Manually create a 4x2 confusion matrix
conf_matrix = np.zeros((len(unique_labels), 2), dtype=int)

# Fill in the confusion matrix based on true labels and predictions
for true_label, pred_label in zip(y_test, predictions):
    conf_matrix[true_label, pred_label] += 1

display_labels = ["M", "F", "M?", "F?"]
sns.heatmap(conf_matrix, annot=True, fmt="d", cmap="Blues", xticklabels=["M", "F"], yticklabels=display_labels)

plt.xlabel("Predicted Label")
plt.ylabel("True Label")
plt.show()

Here is the result:

However I feel like you want this 2x4 instead (As your question was a bit vague) :

import seaborn as sns
import matplotlib.pyplot as plt
import numpy as np
from sklearn.metrics import confusion_matrix

y_test = [0, 1, 1, 1, 2, 2, 3]
predictions = [0, 1, 1, 1, 0, 1, 0]
unique_labels = sorted(set(y_test + predictions))

# Manually create a 2x4 confusion matrix
conf_matrix = np.zeros((len(unique_labels), 2), dtype=int)

# Fill in the confusion matrix based on true labels and predictions
for true_label, pred_label in zip(y_test, predictions):
    conf_matrix[true_label, pred_label] += 1

display_labels = ["M", "F", "M?", "F?"]
sns.heatmap(conf_matrix.T, annot=True, fmt="d", cmap="Blues", xticklabels=display_labels, yticklabels=["M", "F"])

plt.xlabel("True Label")
plt.ylabel("Predicted Label")
plt.show()

Here is the result:

Answer 2

It give you 4 columns because in your y_test you don't have a binary information (not only 0 or 1).

So basically your ground truth seems to be a ground truth for a multiclass classification (int from 0 to 3).

For example, if you replace your 2 and 3 in y_pred by 0 or 1 you will have a 2x2 matrix as you desire.