Logistic Regression on Graphlab Create.

User 327 | 5/28/2014, 3:28:23 PM

Hi, I'd like to know where to download the test dataset of "train-medical-data.csv" in the example of Logistic Regression of Create? Thanks!

Best Regards, Suijian

Comments

User 327 | 5/28/2014, 7:25:26 PM

Thanks Brian, just what to run the example and make clear the data structure of the input file.

-Suijian


User 238 | 5/30/2014, 5:50:16 PM

As a follow up to question on GraphLab Create logistic regression (LR), I ran the logistic regression on a data set of about 14k rows and 96 columns (partitioned as 60% training and 40% test data), the LR returns 0.9319302 as training accuracy. However, the target value of the predicted/test data was in the range that is not -1 and 1 as the training data (I am using GraphLab Create 0.2). The reported value range are real numbers between something like 1 and 7. I would like to know if there is a known error or if there is an API to normalize the result to expected -1 and 1 range (or 0 and 1 for GraphLab Create 0.3)


User 91 | 5/30/2014, 9:54:30 PM

We are glad that you are using Graphlab Create. We are extremely excited for the 0.3 release and would love for you to try it out.

I would like to clarify a few things:

1) In Graphlab 0.3, the input values for target (prediction and training) have been switched to {0,1} instead of {-1,+1}.

2) Secondly, in Graphlab 0.2, the predict function returned the margin (i.e linear predictor) for each observation in the test set. Hence, you didn't get values that were in {-1, +1}. Behind the scenes, the accuracy was computed (in Graphlab 0.2) by converting the margin to the classes in {+1, -1}.

In Graphlab 0.3, we give the user 3 different options for the predict function:

model.predict(testdata, outputtype='probability') # Default model.predict(testdata, outputtype='class') model.predict(testdata, outputtype='margin')

I hope this clarifies your questions. Don't hesitate to follow up with questions!

  • Krishna