8

I am doing Multiclass Classification and applying Logistic regression on it

When i fitted the data by calling

logistic.fit(InputDATA,OutputDATA)

the estimator "logistic " fits the data.

Now when I call logistic.coef_ it prints a 2D array with 4 rows(I HAD FOUR CLASSES) and n columns(one for each feature)

THIS IS WHAT I SAW ON SCIKIT LEARN SITE:

coef_ : array, shape (n_features, ) or (n_targets, n_features) Estimated coefficients for the linear regression problem. If multiple targets are passed during the fit (y 2D), this is a 2D array of shape (n_targets, n_features), while if only one target is passed, this is a 1D array of length n_features.

Now my query is : Why different coefficients are there for different classes as i need only one hypothesis which would predict the output.

pschueller
  • 4,362
  • 2
  • 27
  • 50
Shivam Duggal
  • 317
  • 2
  • 10

2 Answers2

4

As you have a multiclass case (>2 cases) an one-vs-rest strategy is applied. sklearn creates 4 classiefiers, not only 1. Hence you have 4 hypothesis and 4*coefficents.

Note: I have no clue about the logistic regression classifier, but that is how the sklearn SVM work.

fgoettel
  • 351
  • 2
  • 6
  • Thanks a lot.I got it now.Now using the four classifiers, one with max probability forms the output class. – Shivam Duggal Jul 22 '15 at 13:08
  • So, does logistic.predict uses this method for selecting the appropriate classifier?? – Shivam Duggal Jul 22 '15 at 13:09
  • 1
    In the training phase are 4 classifier generated. One seperating A from BCD, one seperating B from ACD, and so on. A new sample (i.e. predict() ) is fed into all four classifiers. The classifier with the highest proability wins. For example your new class is C, then the probability of the C vs. ABD classifier is higher than the other ones. C is the return value of predict(). – fgoettel Jul 22 '15 at 13:55
0

You are getting coefficients for input features depending on L1 or L2 regularization. IF you dont specify L1 or L2, L2 will be assumed by the model. You can use these coefficients for model optimization or feature engineering

ibozkurt79
  • 225
  • 4
  • 7