如何使用逻辑回归训练高度不平衡的数据进行链接预测

Question

This is my first question here in stackoverflow.

I am new to python and I am trying to implement link prediction problem.

I have a list like this:

list_pos = [('alpha', 'beta'),
         ('beta','gama')
         ('alpha','lamda')
         ('gama', 'lamda'),
         ('euphor', 'tuphor')]

And, I am able to generate negative examples of tuple pairs which do not exist previously as follows:

from itertools import combinations
elements = list(set([e for l in list_pos for e in l])) # find all unique elements

complete_list = list(combinations(elements, 2)) # generate all possible combinations

#convert to sets to negate the order

set1 = [set(l) for l in list_pos]
complete_set = [set(l) for l in complete_list]

# find sets in `complete_set` but not in `set1`
list_neg = [list(l) for l in complete_set if l not in set1]

The output is here:

list_neg = 
[['gama', 'tuphor'],
 ['gama', 'alpha'],
 ['gama', 'euphor'],
 ['lamda', 'tuphor'],
 ['alpha', 'tuphor'],
 ['beta', 'tuphor'],
 ['euphor', 'lamda'],
 ['lamda', 'beta'],
 ['euphor', 'alpha'],
 ['euphor', 'beta']]

However, this leads to the following - for 5 positive examples, I have 10 negative examples.

With more items in original list, finally I will end up with a highly unbalanced dataset having lot of negative examples which will effect my model training scores.

My question is - how to train such unbalanced datasets with good accuracy.

For generating my final dataset, I am using the following -

dflp = pd.DataFrame(list_pos, columns=['user1','user2'])
dflp['link'] = 1
dfln = pd.DataFrame(list_neg, columns=['user1','user2'])
dfln['link'] = 0
df_n = pd.concat([dflp, dfln])
df_n.head()

This way I have a dataset suitable for applying logistic regression

Answer 1

If the dataset is large enough, should try to delete some of the negative examples in order to have a balanced dataset.

If the dataset is not large enough you can still delete some of the negative examples and try cross-validation methods like Leave One Out/ JackKnife. These methods to train models are used when the train dataset is small (train dataset < 100 rows)