2

I've the training data set like this:

0.00479616 |  0.0119904 |  0.00483092 |  0.0120773 | 1
0.51213136 |  0.0113404 |  0.02383092 |  -0.012073 | 0
0.10479096 |  -0.011704 |  -0.0453692 |  0.0350773 | 0

The first 4 columns is features of one sample and the last column is its output.

I use scikit this way :

  data = np.array(data)
  lr = linear_model.LogisticRegression(C=10)

  X = data[:,:-1]
  Y = data[:,-1]
  lr.fit(X, Y)

  print lr
  # The output is always 1 or 0, not a probability number.
  print lr.predict(data[0][:-1])

I thought Logistic Regression always should gives a probability number between 0 and 1.

WoooHaaaa
  • 19,732
  • 32
  • 90
  • 138

1 Answers1

14

Use the predict_proba method to get probabilities. predict gives class labels.

>>> lr = LogisticRegression()
>>> X = np.random.randn(3, 4)
>>> y = [1, 0, 0]
>>> lr.fit(X, y)
LogisticRegression(C=1.0, class_weight=None, dual=False, fit_intercept=True,
          intercept_scaling=1, penalty='l2', random_state=None, tol=0.0001)
>>> lr.predict_proba(X[0])
array([[ 0.49197272,  0.50802728]])

(If you had read the documentation, you would have found this out.)

adhg
  • 10,437
  • 12
  • 58
  • 94
Fred Foo
  • 355,277
  • 75
  • 744
  • 836
  • Thanks so much, do you know how to evaluate the quality of the predictions ? The easiest way ... – WoooHaaaa Aug 04 '13 at 08:19
  • @MrROY: in the most recent version of scikit-learn, 0.14a1, there's a function `log_loss` in `sklearn.metrics` which gives the negative log-likelihood of `predict_proba` output. – Fred Foo Aug 04 '13 at 08:20
  • Does X[0] represent predicted event based on the first column, or the overall columns (X has 3 columns here) – user3378649 Apr 21 '14 at 15:12
  • The rows in the `predict_proba` output correspond to rows in `X`. The columns correspond to classes. – Fred Foo Apr 22 '14 at 10:05