Details
Major Description
This is a bug that appears while fitting a Logistic Regression model with `.setStandardization(false)` and `setFitIntercept(false)`. If the data matrix has one column with identical value, the resulting model is not correct. Specifically, the special column will always get a weight of 0, due to the special check inside the code. However, the correct solution, which is unique for L2 logistic regression, usually has non-zero weight.
I use the heart_scale data (https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/binary.html) and manually augmented the data matrix with a column of one (available in the PR). The resulting data is run with reg=1.0, max_iter=1000, tol=1e-9 on the following tools:
- libsvm
- scikit-learn
- sparkml
(Notice libsvm and scikit-learn use a slightly different formulation, so their regularizer is equivalently set to 1/270).
The first two will have an objective value 0.7275 and give a solution vector:
[0.03007516959304916, 0.09054186091216457, 0.09540306114820495, 0.02436266296315414, 0.01739437315700921, -0.0006404006623321454
0.06367837291956932, -0.0589096636263823, 0.1382458934368336, 0.06653302996539669, 0.07988499067852513, 0.1197789052423401, 0.1801661775839843, -0.01248615347419409].
Spark will produce an objective value 0.7278 and give a solution vector:
[0.029917351003921247,0.08993936770232434,0.09458507615360119,0.024920710363734895,0.018259589234194296,5.929247527202199E-4,0.06362198973221662,-0.059307008587031494,0.13886738997128056,0.0678246717525043,0.08062880450385658,0.12084979858539521,0.180460850026883,0.0]
Notice the last element of the weight vector is 0.
A even simpler example is:
import numpy as np from sklearn.datasets import load_svmlight_file from sklearn.linear_model import LogisticRegression x_train = np.array([[1, 1], [0, 1]]) y_train = np.array([1, 0]) model = LogisticRegression(tol=1e-9, C=0.5, max_iter=1000, fit_intercept=False).fit(x_train, y_train) print model.coef_ [[ 0.22478867 -0.02241016]]
The same data trained by the current solver also gives a different result, see the unit test in the PR.
Attachments
Issue Links
- is related to
-
SPARK-13590 Document the behavior of spark.ml logistic regression and AFT survival regression when there are constant features
-
- Resolved
-
- links to
