How to implement a SVM model when the predicted values are pairs of indexes that match?

I am trying to fit a SVM model where my predicted true values are multiindexes that match. The problem is that I dont know how to specify that the multiindixes are the true values.

I cannot use the record linkage classification step because it is not very flexible.

from sklearn.svm import SVC

golden_pairs = filter_tests_new_df[:training_value]
golden_matches_index = golden_pairs[golden_pairs['ev_2'] == 1].index 
# This is a multiindex type

svm = SVC(gamma='auto')
svm.fit(golden_pairs, golden_matches_index) 
# I dont know how to specify that the golden_matches_index are the good matches

# Predict the match status for all record pairs
result_svm = svm.predict(test_pairs[columns_to_keep])

You don't have to specify the index, instead use the generated boolean Series as the labels for the classification.

Here's an example.

# Sample data
data = pd.DataFrame({'a': [1, 2, 3], 
                     'b': [1, 1, 0]})

data
   a  b
0  1  1
1  2  1
2  3  0

# Generating labels
data['b'] == 1
0     True
1     True
2    False
Name: b, dtype: bool

# Can convert them to integer if required
(data['b'] == 1).astype(int)
0    1
1    1
2    0
Name: b, dtype: int64

Based on your code, I think this should do the trick

# Boolean
golden_pairs['ev_2'] == 1

# Integer
(golden_pairs['ev_2'] == 1).astype(int)