[python] How does the class_weight parameter in scikit-learn work?

I am having a lot of trouble understanding how the class_weight parameter in scikit-learn's Logistic Regression operates.

The Situation

I want to use logistic regression to do binary classification on a very unbalanced data set. The classes are labelled 0 (negative) and 1 (positive) and the observed data is in a ratio of about 19:1 with the majority of samples having negative outcome.

First Attempt: Manually Preparing Training Data

I split the data I had into disjoint sets for training and testing (about 80/20). Then I randomly sampled the training data by hand to get training data in different proportions than 19:1; from 2:1 -> 16:1.

I then trained logistic regression on these different training data subsets and plotted recall (= TP/(TP+FN)) as a function of the different training proportions. Of course, the recall was computed on the disjoint TEST samples which had the observed proportions of 19:1. Note, although I trained the different models on different training data, I computed recall for all of them on the same (disjoint) test data.

The results were as expected: the recall was about 60% at 2:1 training proportions and fell off rather fast by the time it got to 16:1. There were several proportions 2:1 -> 6:1 where the recall was decently above 5%.

Second Attempt: Grid Search

Next, I wanted to test different regularization parameters and so I used GridSearchCV and made a grid of several values of the C parameter as well as the class_weight parameter. To translate my n:m proportions of negative:positive training samples into the dictionary language of class_weight I thought that I just specify several dictionaries as follows:

{ 0:0.67, 1:0.33 } #expected 2:1
{ 0:0.75, 1:0.25 } #expected 3:1
{ 0:0.8, 1:0.2 }   #expected 4:1

and I also included None and auto.

This time the results were totally wacked. All my recalls came out tiny (< 0.05) for every value of class_weight except auto. So I can only assume that my understanding of how to set the class_weight dictionary is wrong. Interestingly, the class_weight value of 'auto' in the grid search was around 59% for all values of C, and I guessed it balances to 1:1?

My Questions

1. How do you properly use class_weight to achieve different balances in training data from what you actually give it? Specifically, what dictionary do I pass to class_weight to use n:m proportions of negative:positive training samples?

2. If you pass various class_weight dictionaries to GridSearchCV, during cross-validation will it rebalance the training fold data according to the dictionary but use the true given sample proportions for computing my scoring function on the test fold? This is critical since any metric is only useful to me if it comes from data in the observed proportions.

3. What does the auto value of class_weight do as far as proportions? I read the documentation and I assume "balances the data inversely proportional to their frequency" just means it makes it 1:1. Is this correct? If not, can someone clarify?

The first answer is good for understanding how it works. But I wanted to understand how I should be using it in practice.

SUMMARY

• for moderately imbalanced data WITHOUT noise, there is not much of a difference in applying class weights
• for moderately imbalanced data WITH noise and strongly imbalanced, it is better to apply class weights
• param class_weight="balanced" works decent in the absence of you wanting to optimize manually
• with class_weight="balanced" you capture more true events (higher TRUE recall) but also you are more likely to get false alerts (lower TRUE precision)
  • as a result, the total % TRUE might be higher than actual because of all the false positives
  • AUC might misguide you here if the false alarms are an issue
• no need to change decision threshold to the imbalance %, even for strong imbalance, ok to keep 0.5 (or somewhere around that depending on what you need)

NB

The result might differ when using RF or GBM. sklearn does not have class_weight="balanced" for GBM but lightgbm has LGBMClassifier(is_unbalance=False)

CODE

# scikit-learn==0.21.3
from sklearn import datasets
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import roc_auc_score, classification_report
import numpy as np
import pandas as pd

# case: moderate imbalance
X, y = datasets.make_classification(n_samples=50*15, n_features=5, n_informative=2, n_redundant=0, random_state=1, weights=[0.8]) #,flip_y=0.1,class_sep=0.5)
np.mean(y) # 0.2

LogisticRegression(C=1e9).fit(X,y).predict(X).mean() # 0.184
(LogisticRegression(C=1e9).fit(X,y).predict_proba(X)[:,1]>0.5).mean() # 0.184 => same as first
LogisticRegression(C=1e9,class_weight={0:0.5,1:0.5}).fit(X,y).predict(X).mean() # 0.184 => same as first
LogisticRegression(C=1e9,class_weight={0:2,1:8}).fit(X,y).predict(X).mean() # 0.296 => seems to make things worse?
LogisticRegression(C=1e9,class_weight="balanced").fit(X,y).predict(X).mean() # 0.292 => seems to make things worse?

roc_auc_score(y,LogisticRegression(C=1e9).fit(X,y).predict(X)) # 0.83
roc_auc_score(y,LogisticRegression(C=1e9,class_weight={0:2,1:8}).fit(X,y).predict(X)) # 0.86 => about the same
roc_auc_score(y,LogisticRegression(C=1e9,class_weight="balanced").fit(X,y).predict(X)) # 0.86 => about the same

# case: strong imbalance
X, y = datasets.make_classification(n_samples=50*15, n_features=5, n_informative=2, n_redundant=0, random_state=1, weights=[0.95])
np.mean(y) # 0.06

LogisticRegression(C=1e9).fit(X,y).predict(X).mean() # 0.02
(LogisticRegression(C=1e9).fit(X,y).predict_proba(X)[:,1]>0.5).mean() # 0.02 => same as first
LogisticRegression(C=1e9,class_weight={0:0.5,1:0.5}).fit(X,y).predict(X).mean() # 0.02 => same as first
LogisticRegression(C=1e9,class_weight={0:1,1:20}).fit(X,y).predict(X).mean() # 0.25 => huh??
LogisticRegression(C=1e9,class_weight="balanced").fit(X,y).predict(X).mean() # 0.22 => huh??
(LogisticRegression(C=1e9,class_weight="balanced").fit(X,y).predict_proba(X)[:,1]>0.5).mean() # same as last

roc_auc_score(y,LogisticRegression(C=1e9).fit(X,y).predict(X)) # 0.64
roc_auc_score(y,LogisticRegression(C=1e9,class_weight={0:1,1:20}).fit(X,y).predict(X)) # 0.84 => much better
roc_auc_score(y,LogisticRegression(C=1e9,class_weight="balanced").fit(X,y).predict(X)) # 0.85 => similar to manual
roc_auc_score(y,(LogisticRegression(C=1e9,class_weight="balanced").fit(X,y).predict_proba(X)[:,1]>0.5).astype(int)) # same as last

print(classification_report(y,LogisticRegression(C=1e9).fit(X,y).predict(X)))
pd.crosstab(y,LogisticRegression(C=1e9).fit(X,y).predict(X),margins=True)
pd.crosstab(y,LogisticRegression(C=1e9).fit(X,y).predict(X),margins=True,normalize='index') # few prediced TRUE with only 28% TRUE recall and 86% TRUE precision so 6%*28%~=2%

print(classification_report(y,LogisticRegression(C=1e9,class_weight="balanced").fit(X,y).predict(X)))
pd.crosstab(y,LogisticRegression(C=1e9,class_weight="balanced").fit(X,y).predict(X),margins=True)
pd.crosstab(y,LogisticRegression(C=1e9,class_weight="balanced").fit(X,y).predict(X),margins=True,normalize='index') # 88% TRUE recall but also lot of false positives with only 23% TRUE precision, making total predicted % TRUE > actual % TRUE