Htydjtk

I am not an expert user. I know that I can obtain the confusion matrix, but I would like to obtain a list of the rows that have been classified in a wrong way in order to study them after classification.

On stackoverflow I found this Can I get a list of wrong predictions in SVM score function in scikit-learn but I am not sure to have understood everything.

This is an example code.

# importing necessary libraries

from sklearn import datasets

from sklearn.metrics import confusion_matrix

from sklearn.model_selection import train_test_split



# loading the iris dataset

iris = datasets.load_iris()



# X -> features, y -> label

X = iris.data

y = iris.target



# dividing X, y into train and test data

X_train, X_test, y_train, y_test = train_test_split(X, y, random_state = 0)



# training a linear SVM classifier

from sklearn.svm import SVC

svm_model_linear = SVC(kernel = 'linear', C = 1).fit(X_train, y_train)

svm_predictions = svm_model_linear.predict(X_test)



# model accuracy for X_test  

accuracy = svm_model_linear.score(X_test, y_test)



# creating a confusion matrix

cm = confusion_matrix(y_test, svm_predictions)

To iterate through the rows and to find the wrong ones, the proposed solution is:

predictions = clf.predict(inputs)

for input, prediction, label in zip(inputs, predictions, labels):

  if prediction != label:

    print(input, 'has been classified as ', prediction, 'and should be ', label) 

I didn't understand what is "input"/"inputs". If I adapt this code to my code, like this:

for input, prediction, label in zip (X_test, svm_predictions, y_test):

  if prediction != label:

    print(input, 'has been classified as ', prediction, 'and should be ', label)

I obtain:

[6.  2.7 5.1 1.6] has been classified as  2 and should be  1

Is the row 6 the wrong row? What are the numbers after the 6.? I am asking this because I am using the same code on a dataset that is bigger than this one, so I would like to be sure that I am doing the right things.
I am not posting the other dataset because unfortunately I can't, but the problem there is that I obtained something like this:

  (0, 253)  0.5339655767137572

  (0, 601)  0.27665553856928027

  (0, 1107) 0.7989633757962163 has been classified as  7 and should be  3

  (0, 885)  0.3034934766501018

  (0, 1295) 0.6432561790864061

  (0, 1871) 0.7029318585026516 has been classified as  7 and should be  6

  (0, 1020) 1.0 has been classified as  3 and should be  8

When I count every line of this last output, I obtain the double of the lines of the test set... So I am not sure that I am analysing exactly the wrong list of predicted results…
I hope to have been enough clear.

Welcome to SE:DataScience.

Here [6. 2.7 5.1 1.6] is the feature of the input instance which is classified wrongly. It is one row from your input feature X = iris.data.

The message means: your SVM use the input feature [6. 2.7 5.1 1.6] to predict a label, and it predicts label=2. The ground truth is label=1.

If you want to print the indices of rows that are classified wrongly, you can use

for row_index, (input, prediction, label) in enumerate(zip (X_test, svm_predictions, y_test)):

  if prediction != label:

    print('Row', row_index, 'has been classified as ', prediction, 'and should be ', label)

Welcome.

In addition to what user12075 mentioned, you could do:

indices = np.arange(y.shape[0])

X_train, X_test, y_train, y_test, idx_train, idx_test = train_test_split(X, y, indices, stratify=y, test_size=0.3,

                                                                         random_state=42)

Then,

for input, prediction, label in zip (indices[idx_train], svm_predictions, y_test):

  if prediction != label:

    print(input, 'has been classified as ', prediction, 'and should be ', label)

The following method works for all kinds of classification problem.

Use list comprehension to find all indices of wrong prediction:

indices = [i for i in range(len(y_test)) if y_test[i] != y_pred[i]]

wrong predictions will then be:

wrong_predictions = test_dataframe.iloc[indices,:]

You can also make indices a new column of wrong_predictions, it would be convenient to compare :)