[MADLIB-777] SVM Regression produces different predictions on multiple runs of the same training and test sets.

• Type: Bug
• Status: Closed
• Priority: Critical
• Resolution: Abandoned
• Fix Version/s: v1.9
• Assignee: Rahul Iyer
• Reporter: Srivatsan Ramanujam
• Created: 02/Jul/13 21:58
• Updated: 22/Mar/18 18:46
• Resolved: 22/Mar/18 18:46

Description

I tested this on MADlib 0.7 but I am not sure if this is version specific.

Attaching the training & test tables (combo_svm_train.sql and combo_svm_dev.sql) and the SQL file to train MADlib's SVM Regression and to predict using the trained model on the dev set.

Each time you run the training & prediction, you get wildly different prediction results (the R^2 varies between -0.50 to 0.50 in the several attempts that I ran the model).

Not sure if this is expected behavior or if there is an error I've overlooked. If it is the expected behavior, the models are unusable unless I train the multiple models in parallel and use some sort of voting to minimize the variation. But this seems serious otherwise.