[SPARK-24473] It is no need to clip the predictive value by maxValue and minValue when computing gradient on SVDplusplus model

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Incomplete
• Affects Version/s: 2.3.0
• Component/s: GraphX
• Labels: bulk-closed
• Assignee: Unassigned
• Reporter: JM Cai
• Created: 06/Jun/18 09:12
• Updated: 08/Oct/19 05:41
• Resolved: 08/Oct/19 05:41

Description

I think it is no need to clip the predictive value. It will change the convex loss function to non-convex, which might have a bad influence on convergence.

Other famous recommender systems and original paper also do not include this step.